3
NEMS ARCHITECTURE

OVERVIEW

Chapter 2 described expected uses and desirable inputs and outputs of a National Energy Modeling System (NEMS). In this chapter a general architecture is proposed for the NEMS.

The proposed NEMS should be designed for simulations and analysis relating to the mid-term time horizon, up to about 25 years in the future. While such a modeling system could be used in principle for both shorter and longer term modeling, these kinds of modeling typically have special analytical requirements. For this reason, the committee recommends that EIA not attempt to use a single model for the three kinds of analysis (that is, for short-term analysis, that of up to about 2 years), medium-term analysis, up to roughly 25 years), and long-term analysis, beyond 25 years). In particular, the Department of Energy (DOE) should not select the mid-term model to be used based on the desire to carry out long-term analysis within the same framework.

For short-term modeling, it is usually important to consider supply and demand fluctuations over the course of an annual cycle. Effects of random events--severe weather, strikes, international hostilities--may be of particular interest. Considering adjustments to projected supply and demand disequilibria (based on forecasted starting point derived from the mid-term model) may also be important, for example, when modeling our vulnerability to energy supply disruptions. This report will not consider in greater depth modeling for the short-term time horizons.



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 45
The National Energy Modeling System 3 NEMS ARCHITECTURE OVERVIEW Chapter 2 described expected uses and desirable inputs and outputs of a National Energy Modeling System (NEMS). In this chapter a general architecture is proposed for the NEMS. The proposed NEMS should be designed for simulations and analysis relating to the mid-term time horizon, up to about 25 years in the future. While such a modeling system could be used in principle for both shorter and longer term modeling, these kinds of modeling typically have special analytical requirements. For this reason, the committee recommends that EIA not attempt to use a single model for the three kinds of analysis (that is, for short-term analysis, that of up to about 2 years), medium-term analysis, up to roughly 25 years), and long-term analysis, beyond 25 years). In particular, the Department of Energy (DOE) should not select the mid-term model to be used based on the desire to carry out long-term analysis within the same framework. For short-term modeling, it is usually important to consider supply and demand fluctuations over the course of an annual cycle. Effects of random events--severe weather, strikes, international hostilities--may be of particular interest. Considering adjustments to projected supply and demand disequilibria (based on forecasted starting point derived from the mid-term model) may also be important, for example, when modeling our vulnerability to energy supply disruptions. This report will not consider in greater depth modeling for the short-term time horizons.

OCR for page 45
The National Energy Modeling System For long-term modeling, modeling structures should be much simpler, because in projecting the longer term there are many more unknowns. Long-term modeling systems stress uncertainty about possible outcomes and allow the uncertainty itself to be easily examined. Issues of interest for long-term energy modeling include fuel substitution, physical limitations of natural resources, technological evolution, lifestyle changes, and effects of population growth. Basic concepts for long-term analysis were discussed briefly in Chapter 2, and these also will not be pursued further here. The proposed modeling system for the medium term would consist of modules linked together in a larger analytical system. It should be possible to run these models separately, all together, or in combinations, depending on the analytical needs. This modeling system should be designed primarily to simulate or project energy futures based on assumptions about policies and other driving variables. NEMS should also be useful for some kinds of planning and optimization. In optimization, possible future patterns are compared to objectives encoded in the model. For example, patterns of energy use the modeler deems optimal can be compared with observations about actual rates of energy use and production under various conditions. In what follows, it is assumed that NEMS models will be used mostly for simulations rather than optimizations. NEMS should be structured to project supply and demand equilibria in U.S. energy markets. Prices, which guide the energy system toward a supply-and-demand equilibrium, should be explicitly accommodated in both supply and demand modules (for discussion of general equilibrium analysis, see Ballard and Goulder, 1985; Roger and Goulder, 1984). The modeling system should be able to analyze the impacts of policy options and other variables on the economy, the environment, and the security of energy supplies. Existing regulatory impacts should be represented in the individual modules and the system should allow analysis of contemplated regulatory regimes. In addition, supply and demand modules should allow the examination of non-policy factors (e.g., demographic trends) that, in addition to prices, shape supply and demand. Individual modules would include energy supply and demand models and an interindustry economic growth model brought together in one “integrating system,” and auxiliary models that would generally use outputs from the integrating system once its solution were found. Figure 3-1 provides a simple illustration of the proposed NEMS architecture. The interindustry growth model would take prices of energy and supply and demand-side investments as inputs and project the economic activities resulting from those inputs. Each energy demand model in the integrating system would take as inputs the prices of energy (P) and levels of economic activity (A), and would produce as outputs the quantities of energy utilized (Q) and the demand-side investments associated with that energy use (I). The integrating system's supply and conversion modules would take as inputs the quantities of energy demanded in this system and provide as outputs the prices at which those quantities of energy would be made available. In addition, the supply and conversion models would provide the levels of investments associated with those energy quantities.

OCR for page 45
The National Energy Modeling System Figure 3-1 Simple representation of proposed NEMS architecture. The models within the integrating system would interact with one another through common data files, which would provide information to models and accept it from them, as appropriate. The modules would run sequentially under the guidance of a control module that would seek convergence to an equilibrium. The control module would take as outputs from the other modules values for prices, energy quantities, and investments, and would store them in the common data files. These values would then be passed by the control module back to the individual modules for the next iteration. The control module would include criteria to determine when the system had come sufficiently close to equilibrium. The integrating system would be run iteratively until a solution was found. For any iteration, the integrating system would begin with an initial set of prices and investments over the entire time horizon and would predict demand-side investments and energy quantities demanded over that horizon. These values for energy demand would then be passed to the supply and conversion models, which would in turn provide prices and investments as outputs, which would likely not be consistent with the assumed initial set. The control module would include an algorithm to successively adjust the trajectory of prices and investments to converge toward an equilibrium. Equilibrium would be obtained

OCR for page 45
The National Energy Modeling System when a fixed point was found, that is, a set of price and investment values such that input and output values were identical for each fuel, region, and time. Once an equilibrium solution is found for the integrating system, then outputs from each of its modules could be used in the three auxiliary models, those examining the economic, environmental, and energy security impacts of the given equilibrium. The rationale and for more detail about this proposed NEMS architecture are provided below. MODULAR ARCHITECTURE The proposed NEMS system is described as modular. Modularity in modeling, however, is a relative concept: all modeling systems can be viewed as modular in that equations can be removed and other equations substituted and the model can be run. The modular system intended here refers to a group of modules that, each taking the limited set of inputs from the control module, could be run separately from one another. As the committee envisions it, the defining characteristic of a modular system is the ability to run the modules within a system separately. These modules should have common data files, from which they take inputs to perform calculations and to which their outputs flow. The system should be able to run any individual module without any of the other modules being loaded into the computer. A modular system would allow users to replace modules with alternative versions conferring great flexibility to address highly varied analysis requirements. To implement such a modular system, the EIA must clearly define what data will be passed among the various modules. All such data would be passed through common data files; none would go directly between the component modules, except between the control module and others. EIA must carefully define input and output variables so that they have the same precise definition in all of the modules. Once attention is paid to the interfaces between the modules and the common data files, the internal structure (and possibly even the software) of any module can be independent of any other. Modularity is not a new concept to EIA modeling, as was noted in Chapter 2. Some modularity characterized even the earliest modeling systems of the EIA and Federal Energy Administration. For example, the Project Independence Evaluation System (PIES) and International Energy Evaluation System (IEES) permitted different demand models to be used to calibrate a constant elasticity demand representation and different oil supply curves to be used in the linear programming representation. The current EIA integrated model--the Intermediate Future Forecasting System (IFFS)--is a modular system. The NEMS system that the committee proposes would be a logical extension of the IFFS. There would be no need to develop the NEMS from scratch. Initially it could rely strongly on existing models, adapting them to the new configuration.

OCR for page 45
The National Energy Modeling System The EIA now stresses modularity as an important goal. EIA papers assessing existing and desired modeling systems have emphasized the importance of the modular approach (EIA, 1990a). The committee strongly endorses this emphasis. Advantages of Modular Systems Modular architecture facilitates the decentralized development and maintenance of modeling system components, the modules (Hogan and Weyant, 1983; Cowing and MacFadden, 1984). Thus, once NEMS model inputs, outputs, and module interfaces are well defined, development of NEMS modules could proceed in a decentralized manner under EIA supervision. The ability to substitute one component for another in a modular system allows, among other things, running any combination of modules without running the others (e.g., by using “null modules,” that do nothing and do not change the output values stored in the data files, in place of the full modules). An important related advantage of modular systems is the ability to create and use reduced-form versions of the modules. Reduced-form versions approximate the more detailed modules in translating input variables (e.g., prices for demand or quantities for supply) to output variables (e.g., quantities and investments for demand and prices and investment for supply). Reduced-form versions can be easily substituted for the full modules. They would not be independent models, but simple mathematical structures estimated from the original modules to approximate these modules' full responses. Such reduced-form versions would allow great flexibility in the use of the NEMS. When outputs from particular full modules of the system were not important, these components could be replaced by their reduced-form versions. If many such substitutions were made, the NEMS could be quickly run. Frequent fast runs would be particularly valuable when NEMS were being used for probabilistic analysis. In addition, when many reduced-form versions were used, users could routinely conduct in-depth analyses using the full module of interest without being burdened by the computational costs of running all the full modules. The committee believes that the NEMS could represent crucial relationships in the overall energy system using reduced-form modules, with little deviation in projections from those of the full modular system. Because these simplifications can introduce further error, work needs to be done to assure that the increased variability attributable to the use of reduced-form modules is held to acceptable levels. Modular structure could allow the testing of uncertainty, including that associated with the structure of alternative modules. Simply removing one model and replacing it with another of a different structure would allow testing the uncertainty associated with alternative beliefs about various components of the energy system. Similarly, modularity would allow NEMS to integrate results from external models whenever this were desirable. The committee below recommends that NEMS be developed to run on personal computers if the resulting configuration would not significantly sacrifice the NEMS models'

OCR for page 45
The National Energy Modeling System content. Modular architecture could make this goal easier to achieve inasmuch as only some parts of the entire system would need to be in computer memory at a given time. Disadvantages of Modular Systems. Modular architecture can create coding difficulties not seen in single integrated models that run with one kind of software. The modular structure proposed for the NEMS would rely instead on a control module to guide the operation of the other modules. Off-the-shelf software may not be available and EIA will probably need to develop such a control module. Decentralized development of the NEMS could well lead to few (if any) people understanding the entire model. Thus, few may be capable of interpreting and assessing it as a whole. This is a legitimate concern. The committee believes that EIA should explicitly structure its efforts to avoid this difficulty. In particular, one or more EIA analysts should be charged with overall knowledge of all modules, even those not developed in EIA. Time and effort should also be devoted to ensure that the model as a whole--or larger aspects of it--are understood by those with an appropriate range of strong expertise in modeling and energy-related matters, but not necessarily strong modeling backgrounds. It should also be noted that a modular design places a heavy burden on the architects of the system to think through and specify the requirements of individual modules, and to interact with module-builders in a continuing process of validation and modification. Decentralized development could lead to significant quality control problems in both model development and model maintenance. EIA must take responsibility for maintaining high quality, even for modules developed outside of EIA and for making sure that EIA personnel fully understand them. INTEGRATING MODEL OPERATION The integrating system would take a set of inputs--for example, world oil price, U.S. population growth, and policy and other assumptions--and run the modules until the system converged to an equilibrium. Results from the integrating framework would then be used as inputs to run satellite modules. The solution of the integrating model would typically represent market clearing, that is, supply being equal to demand, for every energy commodity, in every region, at each time. To find a solution, the modules would take as inputs price, quantity, and investment, or some other limited number of vectors from common data files. Once these modules were run, outputs--values for price, quantity, and investment, or other vectors--would be passed back to the common data files. These vectors would be indexed by energy commodity, region, and time. Environmental impact vectors could also be passed to and from the modules in finding a solution.

OCR for page 45
The National Energy Modeling System Figure 3-2 illustrates one convergence to an equilibrium, for energy price and energy quantity. The demand module response is represented by the downward sloping curve, which projects the energy quantity demanded at each energy price. The supply module response is represented by the upward sloping curve, which projects energy supply price for each energy quantity supplied. Equilibrium is represented by the values for price and quantity at which the two curves cross. In Figure 3-2, an initial price is chosen as P0 (for illustration, P0 is chosen far away from the equilibrium price). The price P0 would be read from the common data files by the demand module. The demand module would then project Q0, a value for energy quantity that would be fed into the common data files. The supply module would then read the quantity Q0 from the files and provide a supply price of P'0, which would be sent back to the common data files. The integrating system would be at a solution if P'0 were close enough to P0. If not, the next iteration would use both these prices to calculate a new price estimate, possibly a weighted average of the input and output prices. This new starting point is represented by P1. The process is repeated, giving an output price of P'1. And so the process would continue until input and output price values were sufficiently close. Figure 3-2 illustrates an adjustment process such that energy prices and energy quantities converge to the equilibrium. Yet in other modeled adjustment processes, no equilibrium may be reached, but instead proposed solutions oscillate around the equilibrium point. For example, if the starting point of each iteration were set equal to the output price from the last iteration, the proposed solution could oscillate and would always oscillate for some shapes of the supply-and-demand curve. Additionally, when the number of prices and quantities is greater than those in Figure 3-2, the process of convergence to the equilibrium may be more complex. This complexity may slow the process of convergence or may prevent convergence from being reached. To avoid such problems, EIA should adopt or develop appropriate algorithms for NEMS. For example, by choosing an appropriately weighted average of the initial prices (P0) and the output prices from a given iteration (P'0), convergence can be guaranteed for most supply and demand curves. The appropriate choice of weights also generally greatly decreases the time to convergence. EIA should devote some attention to adapting or developing algorithms to ensure efficient convergence to equilibrium. In Figure 3-2 only a unique equilibrium, or single equilibrium, point exists. Because the supply module response is upward sloping and the demand module response is downward sloping, these two curves can cross only once: the equilibrium is unique. However, if there were many prices and quantities, there might well be more than one equilibrium. If there are multiple equilibria, policy analysis and forecasting becomes especially difficult. With multiple equilibria, the model typically provides no indication which of the several possible system equilibria would be obtained. Therefore, unless the analyst can

OCR for page 45
The National Energy Modeling System Figure 3-2 Illustration of the convergence of price to a supply-demand equilibrium. assess which of the equilibria is most likely to occur, the projections would remain ambiguous. In addition, policy analysis is hampered when multiple equilibria exist. Unless each equilibrium point is examined, the projected impacts of policy measures can be extremely misleading. Figure 3-3 abstractly illustrates the difficulty for the situation in which two outcomes, labeled “Variable 1” and “Variable 2” are being examined. In Figure 3-3, for the first policy measures, the group of points labeled “equilibrium set” could all be equilibria, while for the second policy measures, the group of points labeled “equilibrium set B” could all be equilibria. Note that equilibrium set B is translated downwards and to the left from equilibrium set A. The interpretation is that, whatever

OCR for page 45
The National Energy Modeling System Figure 3-3 Illustration of multiple equilibria. equilibrium point in fact occurs, changing to the second policies will decrease the values of both variables. But a model run does not trace out all the equilibrium points. Rather, in a given run, variables are sequentially calculated until one equilibrium point is found. This process is represented by the crooked lines converging to points A and B. The points A and B would then typically be interpreted as the outcomes occurring with the first and the second policies, respectively. The difference between points A and B would typically be interpreted as the impacts of the policy choices. Notice that point B lies upward and to the right of point A. Thus, if these two points were the only ones that had been determined using the model, it would be natural to conclude that changing to the second policy would increase the values of both variables.

OCR for page 45
The National Energy Modeling System This estimate of policy impacts would be incorrect, as indicated by the shifting of the equilibrium sets. When multiple equilibria exist, therefore, modelers must trace out the set of equilibrium points and examine how the set of points change in response to policy measures. It does not suffice to find only one solution for each configuration of policies. We do not know whether the NEMS will have a unique equilibrium or multiple equilibria. However, EIA personnel must determine which is the case. There are theoretical results that suggest the NEMS might have a unique equilibrium. These results depend upon mathematical properties of the matrices of supply-and-demand derivatives. However, we do not anticipate that all of the assumptions required to prove uniqueness will in fact be valid for the model. For that reason, EIA must explore the model, searching for multiple equilibria, until it is determined that multiple equilibria do exist or until confidence is gained in the idea that the equilibrium is unique. In developing the model, careful attention should also be paid to the hardware to be used in running the model. The committee believes that NEMS should be configured to run on personal computers if such a configuration would not sacrifice the content of the model.1 Modular architecture could make such a configuration easier to achieve. The control module could be designed so to load the modules sequentially into personal computer memory. The control module itself would always remain in memory, while other modules and data files would reside on the hard disk, with only one module at a time in random access memory (RAM). As long as each module were small to be run alone (in conjunction with the control module), then the entire system could be run on a personal computer. For personal computers with sufficient memory, the control module could keep several or all component modules in RAM, running them sequentially, and not expending time or frequent disk swaps. If the NEMS model cannot be adequately structured for use with a personal computer, then it should be configured to operate on a workstation using current software that allows transportability among computers. If the system could be run on personal computers, it could be made widely available, increasing the number of analysts who understood, used, and tested it. Configuring NEMS for workstations would reduce the breadth of its use outside DOE and EIA, but it would still be more accessible than if designed for use only on mainframe computers. The committee believes that such model transportability is important for allowing non-EIA users to understand, critique, and offer ideas for model improvement. For this reason, the committee recommends that NEMS be developed to run on widely available computers unless the resulting configuration would significantly sacrifice the NEMS models' content. 1    For use within EIA, the entire NEMS should perhaps be available on a central file server networked with individual personal computers or workstations. Our recommendation does not preclude that possibility. Such a configuration would allow many analysts to access the same version of the model simultaneously and would allow each access to the most current version, even when the model were being continually revised.

OCR for page 45
The National Energy Modeling System PROPOSED MODULES Figure 3-4 shows the general proposed structure of the NEMS in more detail. Modules shown in double frames indicate models that, to the committee's knowledge, do not exist now within EIA. Separate regionally disaggregated modules would represent various energy-consuming sectors (residential, commercial, industrial, and transportation). These modules would provide as outputs the quantities of the various energy forms demanded for the input delivered prices of energy. Separate modules would also represent various energy supplies (gas, oil, coal, and renewables). These modules would additionally represent U.S. regional transportation of energy, so that supply modules would provide as outputs delivered prices. Single-fuel conversion processes (e.g., refining of crude oil into products) would be represented within the corresponding energy supply model (e.g., the model for oil). These supply modules would take as inputs regional demand and provide as outputs regional prices over time by energy form. Figure 3-4 Proposed NEMS architecture in greater detail.

OCR for page 45
The National Energy Modeling System NEMS COMPARED TO CURRENT DOE MODELING Since the mid-1970s, several large-scale integrated modeling systems of energy supply and demand have been developed at EIA and later in the Office of Policy, Planning and Analysis, for use in recurring publications of energy forecasts and for policy analysis. The modeling systems include the Project Independence Evaluation System (PIES), later renamed the Mid-term Energy Forecasting System (MEFS), the Intermediate Future Forecasting System (IFFS), the Long-Term Energy Analysis Program (LEAP), the Short-Term Integrated Forecasting System (STIFS), and Fossil2 (see Appendix E). One trend exhibited in these models has been for increased modularity and decomposition. Another related trend has been the increased diversity of modeling methodologies employed in the modeling systems. These trends have evolved as system coverage and the complexity of energy issues treated have both increased. The model structure proposed in this report resembles current EIA modeling systems, particularly the IFFS, which is a modular system. The proposed NEMS would represent a logical continuation of EIA model development. NEMS would draw heavily on current EIA approaches and could use revised versions of many current EIA models. There would be some notable differences, however, between NEMS and the IFFS. Reduced-form versions of full modules have not been constructed. The IFFS system cannot run as a complete system on a personal computer, although parts of it can. NEMS would also require some entirely new models, in particular, the interindustry economic growth model and an environmental impacts model. In addition, current models of economic and energy security impacts are not adequate for the proposed system. Finally, IFFS calculations currently clears markets one year at a time, thereby precluding any representation of rational expectations or other “look-ahead” behavior. The Fossil2 model of the Office of Policy, Planning and Analysis, which was used as the integrating framework for the first NES, is constructed using a different strategy. It is not modular. Rather, it includes all supply, demand, and market-clearing equations in one comprehensive dynamic simulation model. Each sector is represented by a set of differential and integral equations that relate the stocks (energy production, transformation, and consumption facilities and entities) and the flows (energy quantities, prices, and information). It would be very difficult, if not impossible, to substitute alternative component modules in Fossil2. For the NES, the various component parts of Fossil2 were calibrated to mimic the EIA Annual Energy Outlook (AEO) reference case to the year 2010. Fossil2 parameters describing sectoral behavior were changed so that projections (at baseline prices) corresponded to the projections in the AEO base case. However, the response of supplies and demands to prices in general was not calibrated to the external models, such as those at EIA. This use of Fossil2 would be quite similar to the proposed NEMS with all (only) reduced-form versions, if Fossil2 were calibrated so that the responses of outputs to inputs

OCR for page 45
The National Energy Modeling System (for example, quantities to prices) were chosen to correspond to the external models. The key difference, however, is that Fossil2 components may well be more complex than NEMS reduced-form versions. Such greater complexity would provide greater richness of detail and more policy “handles” for analysts, but would make calibration more difficult for the NEMS. Nevertheless, NEMS reduced-form versions could be exactly like the components of Fossil2, if Fossil2 components were deemed the best reduced-form versions of the external models. Fossil2 components then could be used in NEMS if it were possible to separate them in the form of component modules. The Fossil2 model differs from the proposed NEMS in that its structure is explicitly calculated on an annual basis: one year is calculated, then the next, and so on. Given such a structure it is easy to incorporate myopic expectations and adaptive expectations. However, to incorporate rational expectations, the entire model must be run over and over, using the output prices from any run as input prices for the next run. While in principle such an approach is feasible, in practice it would be very time consuming and probably not routinely pursued. The Fossil2 model also differs from the proposed NEMS in its market-clearing mechanism. Prices move over time in response to differences between supply and demand, so that in any given year prices might not cause markets to clear. To minimize the modeling difficulties such nonclearing might cause, Fossil2 is recalculated four times for each year. While four interactions is probably sufficient to bring each year's supply and demand into close approximation, it cannot be guaranteed that markets would clear. NEMS TREATMENT OF CONCEPTUAL ISSUES NEMS architecture should incorporate conceptual issues that are significant to debates about energy market operations. This section considers how the proposed NEMS could address such conceptual issues. Market Disequilibrium NEMS architecture should be based on the idea that energy markets will clear: energy prices will take values such that supply is equal to demand for each energy form (Weyant, 1985). One question is the time it takes markets to clear, given whatever regulatory or other rules exist. The concept of equilibrium is less useful if markets do not clear within the time period modeled, for example, one year. The equilibrium price model assumes that instantaneous adjustments occur; it does not consider the actual time it may take real markets to clear. In some cases, market clearing occurs over a long period as when there are government-imposed price controls and nonmarket allocations of energy forms. Government interventions can create market disequilibria that last for years, if not decades. The NEMS modeling system could be used in a simple way to examine non-market-clearing associated with price controls. Rather than require the price-controlled market to converge to equilibrium, the model user could set product price-say retail price

OCR for page 45
The National Energy Modeling System for gasoline--to the government controlled price. The test for convergence could be modified to allow the price-controlled product market to remain out of equilibrium while other markets converged to supply-demand equilibria (equilibria likely to differ in many markets from those that would obtain without the price controls). Another important question is the market-clearing mechanism. Do markets clear based on fluctuations in price? With other clearing mechanisms (fluctuations in production, for example) the economy still has competitive markets that clear, but these markets have different characteristics and outcomes. Instead of moving to the equilibrium point given by conventional supply-and-demand curves, the markets may clear by shifts in the curves themselves. For example, when there are government imposed price controls on gasoline, the length of the gasoline lines could be expected to adjust to clear markets. In an aggregate model, such adjustments in waiting times could be seen as shifts in the demand functions for gasoline. Such adjustments could be incorporated in the modeling system if it were known what shifts might occur. The success of the equilibrium price-model in predicting politically feasible outcomes depends on relatively large responses on either the supply or the demand side of the market to changes in price. If the responses (elasticities) are both very small, price increases (e.g., in oil) may cause large real-income changes (people cannot reduce their oil bills by buying less), and the public is politically unlikely to let the market work. Thus, the model user must consider political infeasibility of some possible equilibria. Uncertainty As discussed previously, forecasting and policy analysis require understanding the degree of uncertainty in model outputs. As discussed in Chapter 2, careful analysis of uncertainty requires a large number of model runs, using not only the standard modules, but also various combinations of alternative modules. Unless some method is used to shorten the time this analysis required, it is so time-consuming in practice that it is seldom conducted. One requirement for the NEMS, therefore, to analyze uncertainty, is the ability to run the model very quickly. One method of doing so relies on the use of reduced-form versions that can each be run quickly, to rapid system convergence to an equilibrium. Such analysis would be conducted in two stages. In stage one, the uncertainties associated with each individual module would be characterized by exercizing that module. Two elements of uncertainty are particularly important for projections: uncertainty about the projections based on a particular set of inputs, and uncertainty about actual values of the input variables. These uncertainties could be used to construct a stochastic version of the reduced-form version for a given sector. This simplification could be achieved by imposing a probability distribution on the parameters used in the reduced-form version to correspond with the probability distribution associated with the full module. Once the stochastic reduced-form versions were developed, the overall system could be configured to run based entirely on reduced-form versions. Monte Carlo simulation techniques could be used to randomly select combinations of the various parameter values

OCR for page 45
The National Energy Modeling System for each of the reduced-form versions. Then the entire model could be run until it converged to an equilibrium. This process of randomly selecting parameters and running the overall model could be reiterated until sufficient data were obtained. The advantage of using reduced-form versions for such a process is primarily that of speed. The greatly reduced turnaround time that the use of reduced-form modules permits may bring such Monte Carlo simulations to manageable proportions. In addition, the use of reduced-form versions may make it easier to incorporate purely subjective probability distributions for the underlying parameters. This ability might encourage researchers to avoid underestimating the uncertainty associated with the underlying models and therefore to avoid underestimating the uncertainty associated with the system as a whole. Contingent Strategies As discussed previously, many private and public policies cannot be regarded as predetermined, but rather should be seen as responsive to the immediate, externally imposed situation. For example, were it not for the rapid oil price increases in 1973–74, Presidents Ford and Carter probably would not have initiated the energy policy proposals of that period nor would Congress have passed legislation such as that imposing corporate average fuel economy (CAFE) for new cars. Thus, energy demand and supply functions for energy were themselves shaped by unexpected events. Similarly, the U.S. Synthetic Fuels Corporation could be seen as having pursued contingent strategies in attempting to create energy technologies that could be more rapidly developed by the private or public sector, in case changing energy situations created the need. As it turned out, evidence mounted that energy prices would not rise as rapidly as previously expected and that the costs of synthetic fuels would be higher than some had expected, and the entire effort was disbanded. Such contingent strategies are conceptually important in modeling the uncertainty in the energy system. In a deterministic forecast, contingent strategies can be treated exactly like deterministic policy choices: they are adopted or not. Thus, for deterministic modeling there is nothing special about contingent strategies. The essence of contingent strategies is that policy rules themselves may depend on features of the energy system. For example, if energy prices go above a particular point, then energy conservation rules may be imposed. If energy prices remain below another level, or are projected to continue below some level, then such policies may be repealed. In a stochastic modeling environment, contingent strategies will be governed by one set of rules in one state of the world and by another in a different world state. Projecting such endogenous policy changes is particularly difficult and not always relevant. However, if they can be predicted with some probability, then such contingent strategies can be encoded into NEMS. Within the relevant modules, there would simply be logical statements that test whether the state of the system is such that new rules would be triggered or old rules eliminated. If so, then the relevant equations in the modules would be adjusted. Thus, the conceptual difficulty with analysis of contingent strategies is their clear specification in the first place.

OCR for page 45
The National Energy Modeling System There would also be practical difficulties in addressing contingent strategies. It will be much more difficult, if not impossible, to use reduced-form versions of modules that accommodated contingent strategies. However, reduced-form versions could still be used for other modules. Clearly, the analysis and computing time, but especially computing time, for examining contingent strategies may be extremely large. The Formation of Expectations Yet another question, related to that of disequilibrium, is how models should represent expectations of future prices (Muth, 1981a, b; Lucas and Sargent, 1981). Several alternative assumptions are possible. The simplest is “myopic expectations,” that individuals in the market assume future prices will remain the same as current prices (Muth, 1981a, b; Lucas and Sargent, 1981). In a stochastic setting, “myopic expectations” are calculated using the random-walk model that has been highly successful in forecasting financial markets. Alternatively, the “extrapolative expectations” hypothesis assumes that individuals extrapolate recent trends in prices to forecast future prices. Individuals do adjust, but not instantly, to changing information, and a random walk characterizes the trend lines of energy variables. Finally, individuals assumed to have “rational” expectations forecast prices and other economic conditions consistently with the model. When economic actors behave “rationally,” rather than “adaptively” they do not gradually change their behavior in response to new information or different circumstances but adopt new decision rules as information or circumstances change. Decision makers are also assumed to look to the future-- to their expectations-- rather than just to the past--their experiences-- when making decisions. Assuming rational expectations, if the model is deterministic, then individuals forecast prices correctly; if the model is probabilistic, then individuals forecast prices using the same probability distribution about future prices that results from the model. The assumption of rational expectations emphasizes the important point that people learn from their pasts and try not to repeat mistakes. It reflects the observation that behavior sometimes does change quickly and discontinuously. Because expectations about the future are strongly conditioned by our experience, the real question becomes the extent to which our behavior is rational or adaptive. Learning theory researchers have found that people often make systematic mistakes and that they take time to move from one mode of behavior to another. Thus neither theory perfectly predicts actual human behavior. There is a further conceptual distinction between rational expectations and adaptive expectations approaches when incorporated into policy relevant models. Rational expectations, or self-consistent expectations, when incorporated in models, embodies the basic concept that government policymakers cannot know the future better than private actors. Use of rational expectations concepts in NEMS would be consistent with a belief that policymakers cannot be assumed to have such privileged access to information about the future. The use of adaptive or myopic expectations implies that government planning can always improve on private decision, because the government can be assumed to know the future better than private actors.

OCR for page 45
The National Energy Modeling System It is currently an open debate as to which assumptions about expectations most accurately, or most inaccurately, describe the world. And a number of energy policy debates turn on disagreements as to whether governmental regulation can improve on private decision-making from the perspective of the individual actors. Thus the appropriate description of expectations will not be easily resolved. For that reason it is important that the modeling system be capable of addressing each of the assumptions about expectations formation. These assumptions can influence model projections including assessments of particular policy options. Unless the process of expectation formation is considered explicitly by NEMS, it will be difficult to assess the system-wide consequences of programs designed to educate people about energy issues or of behavior shifts occasioned by changes in beliefs. If people systematically underestimated future energy prices and therefore systematically bought more energy-intensive capital equipment than they would otherwise, this belief, if identified, might lead to a policy of advertising that energy prices are likely to increase. In a world of myopic expectations, such a strategy might significantly change behavior. In a world of rational expectations, such a strategy would have little or no effect on behavior, since the assumption of rational expectations is contrary to people systematically understating future conditions. A less obvious example illustrates the feedbacks that may depend on expectation formation. Assume that incentives are proposed to increase the efficiency of newly purchased cars beginning in five years. Such a policy might significantly reduce consumption five or ten years from its implementation. But the consumption reduction might reduce the price of gasoline and thus make it more cost-effective for consumers to buy less fuel-efficient cars in the first few years. If we assume rational expectations, then the model will show that consumers will buy less efficient cars in early years and will therefore consume more gasoline during the first few years the policy is in place. Even over a longer period, the earlier purchases of inefficient cars would increase gasoline use compared to the case if there were no such anticipation of a price drop. If the model assumed myopic or extrapolative expectations, then it would show consumers in the first few years as not anticipating the price reduction and thus not changing their automobile purchases. These models would show that in all years the policy would reduce gasoline use or leave it unchanged. Thus, assumptions about expectations formation can change the magnitude and even the direction of overall impacts of some policy impacts. If either myopic or extrapolative expectations are believed to best describe world, then NEMS can be run on a year-by-year basis and price data can be based on either the current price or the current and previous prices. However, if rational expectations formation is assumed and a deterministic model used, then NEMS must be solved for all years simultaneously, since supplies in any year will depend on prices projected for future years as well as current and past prices. In principle, iteration based upon all prices would be similar to iteration based upon only the current prices. However, the greater dimensionality of the problem could be expected to increase the time until convergence is obtained, perhaps greatly. In such a case, it would be important to devote attention to computationally efficient algorithms.

OCR for page 45
The National Energy Modeling System For a probabilistic model, however, the process would involve an immensely more complex, time-consuming endeavor. In principle, one would start with a probability distribution of prices, associated with the full range of uncertainty. This distribution would lead to one vector of actual prices for each possible outcome. Finding this one vector would require the iterative processes discussed above, and the process would have to be repeated for each resolution of the uncertainty. The result would be a probability distribution of prices over time. If this distribution were consistent with the starting distribution of prices, then the solution would be a “rational expectations” equilibrium. However, if the distributions were not consistent, then the starting distribution would be adjusted and the process repeated. This sequence would be continued until convergence were reached. While in principle rational expectations equilibria could be found in this way for a probabilistic model, in practice the process would be so time-consuming as to be entirely impractical unless better algorithms are developed and computer speeds increased above those currently available. Thus, the NEMS will not be used routinely to examine probabilistic rational expectations equilibria, even though rational expectations concepts are logically most consistent only in a stochastic environment. Yet it is important that the model capture as well as possible the effects on behavior of changing information, particularly information changing as a result of contemplated policy actions. Environmental Constraints The NEMS is likely to be used to examine the economic, environmental, and security implications of various types of environmental constraints. Historically, the instruments for environmental regulation were primarily technology-specific requirements relating to the use of so-called “best available control technology.” Such rules set limitations on the technologies that could be adopted in the regulated sectors. Simulations of the effects of such restrictions could readily be incorporated into the proposed system by modifying descriptions of technologies underlying the various supply and conversion models. More recently, however, more market-related policies have been adopted, such as the marketable-permit system for acid rain precursors. Such regulations impose system-wide or sector-wide constraints on the amount of residuals emitted. System-wide or sector-wide constraints can also be incorporated into NEMS, by bringing some parts of the environmental impacts module into the integrating system. A per-unit emissions price (e.g., a price per ton of airborne sulfur) could be associated with discharge of the measured residual. Total discharge could be calculated at each iteration based on the economic incentives associated with the estimated emission price. Equilibrium would obtain when the total emissions “supplied” by the system are identical to the maximum emissions allowed. When these two quantities differ, the emission price could be adjusted from iteration to iteration in the same way other prices are adjusted. Figure 3-5, illustrates the convergence process for system-wide environmental constraints. This figure is similar to Figure 3-2, but in place of the demand module response is the maximum total emission, set by the regulations. The emission “supply” response is represented by the downward sloping curve, which projects total emission supply

OCR for page 45
The National Energy Modeling System price for each possible total emission level. This supply price would be based on the estimated behavior of the industries modeled. Equilibrium would be represented by the emission price and emission quantity at which the two lines cross. In Figure 3-5, an initial price is chosen, P0, far from the equilibrium price. The maximum total emission remains at QMax. The modules projecting emissions would then read the price P0 and provide an emission estimate of Q0. The integrating system would be at a solution if Q0 were close enough to QMax. If Q0 exceeded QMax, the next iteration would be based on a higher price, and conversely, if Q0 were smaller than QMax, the next iteration would be based on a lower emission price. This new starting point is represented in Figure 3-5 by P1. The process is repeated until projected and maximum emission are sufficiently close. OPERATIONAL ISSUES FOR NEMS DEVELOPMENT Several operational issues should be given careful attention in constructing a modular system. Such attention should be given to the tradeoffs among model detail, needs for data to support the model, computing hardware, and software requirements. It is useful for each module to be scaled to about the same size and detail, so that no single module dominates the choice of computer hardware and limits the usability of the overall system. However, the EIA models include at least one--the coal model--that is far more detailed and larger than useful for the NEMS. Such models will need to be greatly streamlined and simplified. Other models may need expansion to provide the detail appropriate for the system. Although a modular system could be developed through decentralized teams, in practice full decentralization is not feasible. Rather, constraints should be imposed beyond those implied by the interfaces. For example, the NEMS office may specify that a given module must not require more than a certain amount of comuter memory, to allow the entire system to be operated on a personal computer. EIA personnel could be expected to examine models for quality assurance. Particular coordination would be required to assure that modules all use common data bases, common economic assumptions, and regional, product, and time definitions, as appropriate. The time sequence of efforts will be important, because it would be unrealistic to expect all NEMS modules to be developed during the first year of NEMS efforts. Perhaps construction of the integrating framework and the control module should be given the highest priority. Early attention could be paid to reducing the size of any unnecessarily large model that would be included within the integrating system. The EIA should build the integrating framework and control module from scratch, after determining the appropriate regional, temporal, and product definitions. However, development of the other modules probably should initially rely heavily upon models already in existence in EIA or elsewhere. Such development might involve bridge or conversion systems to link existing modules into the new integrating framework. It might involve recalibration of modules to the new regional, temporal, and product structure. Or it might involve redevelopment of existing models, using existing conceptual structures, empirical evidence,

OCR for page 45
The National Energy Modeling System Figure 3-5 Modeling a system-wide environmental constraint. and data, but entirely new computer code. Over time, it is envisioned that new modules would be developed and would supplant these modules. Several issues are particularly important for redeveloping existing modules. EIA should incorporate the capability of projecting key system-wide environmental impacts into each of the models as they are redeveloped. In particular, emissions of various air pollutants should be estimated for each of the modules within the system. While modules are being constructed, EIA should give attention to building in to the modules explicit representation of factors that might be modified by specific policy actions. Examples of policy actions that the committee believes to be important include those associated with taxes, environmental constraints, conservation incentives and regulations, and new technology development.

OCR for page 45
The National Energy Modeling System To use reduced-form versions, EIA would have to develop the appropriate methodologies and software for estimating response surfaces. While other researchers have developed methods, EIA personnel or contractors would need to further develop these approaches or to contract for their further development. Because reduced-form versions would greatly facilitate NEMS use for quick-turnaround analyses, the development of the methods and software to estimate response surfaces should be given fairly high priority. RECOMMENDATIONS. The NEMS should be designed to be modular in structure, so to be flexible in its use and to readily accommodate the substitution of alternative models (modules) in the system. A single model should not be used for short-term, medium-term, and long-term analysis. Similarly, the DOE should not select the model to be used for medium-term analysis based on the desire to conduct long-run analysis through the same framework. The EIA should attempt to acquire an existing interindustry growth model and should not attempt to develop one itself. Development of the NEMS can rely extensively on existing models. However, several new models should be developed or acquired from external sources: beyond the interindustry economic growth model, these should include an environmental impacts model and a renewable energy supply and conversion model. Demand models need to be modified to enable a broader range of policy analysis than is possible with the current EIA models. A reduced-form version of each module should be developed and integrated in to the NEMS. Such versions should approximate the response surface of the full modules. For typical policy analyses, the full module would be used for the sector examined and reduced-form versions would be used for other sectors. Reduced-form versions could also be used in uncertainty analyses, tests of the integrated set of modules, and quick-turnaround policy analyses. The NEMS should be configured to run on personal computers or workstations, unless such hardware constraints would entail significant loss of the capabilities envisioned for NEMS. One or more EIA analysts should be charged with the general knowledge of all modules, even those not developed at EIA. In developing the environmental module, initial efforts should be focused on quantifying direct and indirect air emissions. If resources are not sufficient for this

OCR for page 45
The National Energy Modeling System task, then EIA should focus attention first on the greenhouse gas emissions associated with the extraction, production, transportation, and use of energy. The NEMS should capture the effects on behavior of changing information, particularly information changing as a result of contemplated policy actions. Model users should be able to include alternative assumptions about the formation of expectations, including those defined theoretically as myopic, adaptive, and rational expectations. The NEMS should provide carefully envisioned graphics and report writers to provide routine graphical and numerical outputs of the types normally found helpful for analytical purposes. A major effort is needed to collect more extensive data and information on the U.S. energy system, especially on end use. In such an effort, DOE and EIA should consider the following points: Essential to bottom-up demand modeling is knowledge of underlying activities: housing, commercial buildings, industrial production, and transportation. To obtain such information, EIA needs to improve its link to other data-gathering entities. Where behavioral information is inadequate, EIA or other DOE offices should help generate interest in obtaining it, by soliciting research, holding workshops, or stimulating other agencies to sponsor research. EIA should devote some resources to sponsoring research in ayuiathis area.