Electricity Demand Forecasts ShoreENERGY The Energy, Economic, and Sustainability Program of the Franklin P. Perdue of Business at Salisbury University Dan M. Ervin Ph. D. Director E. Tylor Claggett Ph. D., CFA Associate Director
1 Type of Forecast Electricity Demand Forecasts Load forecasts can be categorized in many different ways. The report uses the length of the forecast as the initial discriminator. Each forecast, short-term, medium-term, and long-term, has different challenges and characteristics which require different inputs and or techniques. Each type of forecast faces similar risks; however, the importance of each risk factor can be very different. 1.1 Short-term forecasts In general, short-term forecasts are for any time period of one year or less. This type of forecast is not greatly affected by regulatory or technological risk. However, a sudden loss of a large industrial customer could have large impact on this type of forecast. In addition, unusual circumstances may lead to these risks affecting the validity of short forecasts. For example, Maryland may find power suppliers reluctant to provide service as re-regulation discussions increase in the State legislative process. 1.2 Medium-term Forecasts Medium-term forecasts are for one to five year time periods. These forecasts can be used to determine the need for generation assets requiring a short permitting and construction schedule such as peaking units. They are not useful for determining baseload generation requirements. Other five-year considerations must address the possibility of technological change. These changes occur at a quick pace and could affect the medium-term forecast. Conversely, the construction period for modern baseload assets is typically much longer than five years. 1.3 Long-term Forecasts Long-term forecasts are for periods greater than five years. These forecasts are useful to a variety of interested parties. An important use of this type of forecast is resource planning. In deregulated states, generation companies can use these models for resource planning activities. For assets that are still regulated, i.e. transmission assets, this type of forecast is needed for proper regulation and planning. 2 Forecasting Methods 2.1 Econometric Models A large number of regression models is available for use in forecasting. The most suitable model depends on the characteristics of the data. Furthermore, regression models can be combined with time series models to produce more comprehensive forecasts. This report does not provide an overview of econometric models.
A large number of regression models is available for use in forecasting. The most suitable model depends on the characteristics of the data. The sophistication of the regression model can vary from simplistic to very complex. The software producing the forecast can range from the ubiquitous MS Excel to custom, proprietary programs. Regression models can be combined with time series models to produce more comprehensive forecasts. These models, panel data models, have an increased complexity when compared to regression models. 2.2 End-User Models End-User models are used in conjunction with many different forecasting methods. These models provide insight into the power requirements for different end-user categories. For example, a residential end-user model would examine a prototypical household and the electrical requires for that particular class of households. Different types of households can be defined such as apartments, town homes and single-family homes or by living space square-footages, etc. 2.3 Monte Carlo Simulation Monte Carlo Simulation is a robust method with many applications. Each type of forecast can be simulated. This technique can include the possible variations and associated probabilities (riskiness) of each input in an explicit manner. In addition to defining each variable s distribution, the correlations between variables can be included in the model. Advanced models can incorporate dynamic effects in the simulation. 2.4 Sensitivity Analysis Each forecasting method can and should include a sensitivity analysis. The sensitivity analysis identifies the variable with the largest impact on the forecast. The analyst can use this information to explore the characteristics of this variable. It is possible that an in depth understand of the variable(s) contributing the most risk to the forecast will allow the forecast users to make appropriate contingent plans. Develop a more accurate estimate of this variable if required. 2.5 Scenario Analysis Scenario analysis is much like a simulation analysis. The model is based on some number of scenarios, usually 3 or greater, and the variables are combined in a way that will produce several outcomes. A typical scenario analysis will produce a worst, expected, and best case. 3 Input Assumptions
The inputs for the forecast will depend on the type of forecast and customer type or rate class. The drivers of a long-term forecast for industrial customers will have different inputs when compared to a short-term residential forecast. Another consideration is the level of detail for each input variable. A particular service area may grow at a rate faster or slower than the state. Usually it is cost effective to develop data based on a smaller area such as a county or sub-station. The customer type influences the input information. Economic data is important for each customer type; however, its impact is normally stronger in an industrial forecast when compared to a residential forecast. Weather is more important to residential than industrial forecasts. 4 How do you model impacts from demand side programs? To include uncertainties such as demand response, conservation, and energy efficiency programs in a forecast requires a high level of skill and judgment. Two basic questions must be answered. When will the program(s) impact the demand and to what extent will they change consumer behavior; thereby changing electricity usage. Each question is difficult to answer. Certainly power price and the elasticity of electricity will have an impact on these two issues. Again, each type of forecast will be affected differently by the answers to the previous two questions. If the program introduction period is simultaneous with the short-term forecast period, there could be a significant impact on the actual outcome. Otherwise, the short-term forecast should not be influenced. If the price of power continues to grow at substantial rates, consumers will be influenced by energy efficiency regardless of programs introduced by the utilities or the states. If time-of-day rates, with transparent prices, are introduced, consumers will alter their behavior by shifting as much power consumption as possible to times with the lowest cost. These types of behavior modifications could affect forecast accuracy and will present a challenge to the forecaster. These challenges can be quantified using sensitivity analysis. Perhaps market testing these programs may provide insights as to how these programs or any demand side programs will affect forecast outcomes. Long-term forecasts face all of the risks faced by medium-term forecast. However, the difficulty of including these programs becomes more and more difficult as the forecast horizon increases. 5 Weather/Climate Factors Weather is always an important variable in any electrical energy use forecast with the possible exception of industrial forecasts. Therefore, weather and climate factors must be included in each type of forecast. The level of detail will change for each.
As any forecaster knows, actual electrical loads are affected extensively by temperature and other weather factors. Traditionally, historical heating and cooling degree days (HDD and CDD) have been used to forecast the electrical load. Simulations can easily include this affect. However, wind, cloud cover and rain can reduce demand in the summer and increase it in the winter. Perhaps this information too should be modeled. The short-term forecast may include monthly forecasts. The seasonality associated with residential electrical usage is well documented. As the length of the forecast increases, the uncertainty of the weather effects increases. The number of methods to include weather factors also increases and no one way seems to dominate the others. For example, the forecaster may use average HDDs and CDDs or the sum of HDDs and CDDs. Different bases for HDDs and CDDs may provide more accurate forecasts. But, once again, sensitivity analysis may provide insight as to the best method. Finally, simulations provide powerful tools for including the variability inherent in weather statistics. Trends can be added to provide a more dynamic element to the forecast. 6 Model Accuracy Model accuracy is an important feature of any forecast. As forecasters know it is difficult to include all of the variables and the correct statistical characteristics of each included variable. An experienced forecaster develops judgment skills that play an important role in their work. Feedback from forecasts provides much of this learning. Important characteristics of forecasts can be learned as they are compared to actual demand data.