The authors [PS29] discuss a power load forecasting system based on a temporal difference (TD) method. The temporal difference method is a class of statistical learning procedure specialized for future prediction. The conventional BP has been successfully applied to power load forecasting in a one-to-one single-step prediction manner. The authors adopt a method which can gradually improve its prediction accuracy through the time evolution. In addition, a modified temporal difference (MTD) method which uses a different error function is evaluated and compared with the original one.
The authors [PS30] present a method of short term load forecasting using an ANN. A three layered feedforward adaptive ANN, trained by BP, is used. This method is applied to real data from a power system and comparative results with other methods are given.
The authors [PS31] describe an ANN approach to improve short term forecasts of electricity demand. This network is based on the nonstatistical neural paradigm, BP, which is found to be effective for forecasting of electrical load. The load is decomposed into a daily pattern reflecting the difference in activity level during the day, a weekly pattern representing the day-of-the week effect on load, a trend component concerning the seasonal growth and a weather component reflecting the deviations in load due to weather fluctuations. The performance of this network has been compared with some commonly used conventional smoothing methods, and stochastic methods in order to demonstrate the superiority of this approach.
The paper [PS32] presents a method of changing a topological ANN to forecast the load of a power system. The model is almost an all-round reflection of various factors which affect the changing of load. The data window which the ANN needs to learn the BP algorithm is the shortest. Through man-machine dialog, the load forecasting of various forecasting terms can be flexibly realized. Calculations show that the method is efficient. The forecasting accuracy is more accurate than that of conventional methods. It has satisfactory convergency and high computing speed, which is suitable for online application.
The paper [PS33] addresses training data sensitivity problem of ANN-based power system load forecasting. A crucial problem with the ANN-based load forecasting is that its forecasting performance is significantly affected by the selection of training data used to calculate the network weights. The inherent shortcoming of this approach is verified through a typical example presented in this paper. Test results show that the short-term load forecasting error is very sensitive to the amplitude of the noise signal which is added to a portion of the training data. The presented test cases approximately simulate the load conditions during abrupt weather changing periods. Possible strategies to remedy this problem are also discussed in the paper.
The paper [PS34] presents a short-term load forecasting technique for summer using an ANN. The purpose of this study is to forecast accurately daily peak load for a target period using actual data from the same period of the previous several years as training data. This paper describes two methods. In one method, the actual data of each year for the several years earlier are used for each ANN. The other method uses the collective data of several years for the training of the ANN. With the proposed method, the mean absolute forecasting error was below 2%.
The paper [PS35] presents an ANN based model for the calculation of next day's load forecasts. The model's most significant aspects fall into the following two areas: training process and selection of the input variables. Insights gained during the development of the model regarding the choice of the input variables, and their transformations, the design of the ANN structure, the selection of the training cases and the training process itself are described in the paper. The new model has been tested under a wide variety of conditions and it is shown in this paper to produce excellent results. Comparison results between an existing regression-based model that is currently in production use and the ANN model are very encouraging. The ANN model consistently outperforms the existing model in terms of both average errors over a long period of time and number of "large" errors. Conclusions reached from this development are sufficiently general to be used by other electric power utilities.
The paper [PS36] presents the development and implementation of an ANN based short-term system load forecasting model for the energy control center of the Pacific Gas and Electric Company (PG&E). Insights gained during the development of the model regarding the choice of the input variables and their transformations, the design of the ANN structure, the selection of the training cases and the training process itself are described in the paper. Attention was paid to model accurately special events, such as holidays, heat-waves, cold snaps and other conditions that disturb the normal pattern of the load. The significant impact of special events on the model's performance was established through testing of an existing load forecasting package that is in production use. The new model has been tested under a wide variety of conditions and it is shown in this paper to produce excellent results. Comparison results between the existing. regression based model and the ANN model are very encouraging. The ANN model consistently outperforms the existing model in terms of both average errors over a long period of time and number of 'large' errors. The ANN model has also been integrated with PG&E's Energy Management System (EMS). It is envisioned that the ANN model will eventually substitute the existing model to support the Company's real-time operations. In the interim both models will be available for production use
This paper [PS37] presents an approach for short term load forecasting using a diagonal recurrent ANN with an adaptive learning rate. The fully connected recurrent ANN (FRNN), where all neurons are coupled to one another, is difficult to train and to converge in a short time. The DRNN is a modified model of FRNN. It requires fewer weights than FRNN and rapid convergence has been demonstrated. A dynamic BP algorithm coupled with an adaptive learning rate guarantees even faster convergence. To consider the effect of seasonal load variation on the accuracy of the proposed forecasting model, forecasting accuracy is evaluated throughout a whole year. Simulation results show that the forecast accuracy is improved.
The method proposed [PS38] is aimed to accurately forecast the daily peak loads for a target period using the actual data for the same period of the preceding year as training data. This paper describes the peak load forecasts during summer and winter respectively, since the seasons affect the peak load differently. The ANN performance realized a standard deviation error of 2.3% in summer and 1.6% in winter.
In the paper [PS39], the effectiveness of an ANN approach to short-term load forecasting in power systems is investigated. Examples demonstrate the learning ability of an ANN in predicting the peak load of the day by using different preprocessing approaches and by exploiting different input patterns to observe the possible correlation between historical load and temperatures. A number of ANNs have been demonstrated with emphasis given to their practical implementation for power system control and planning purposes. The network is trained on actual power utility load data using a backpropagation algorithm. The prospects for applying a combined solution using ANNs and expert systems, called the expert network, is also discussed. It may give a more complete solution to the forecasting problem than either system alone can provide.
An ANN approach is proposed [PS40] for one-week ahead load forecasting. This approach uses a linear adaptive neuron or adaptive linear combiner called Adaline. An energy spectrum is used to analyze the periodic components in a load sequence. The load sequence mainly consists of three components: base load component, and low and high frequency load components. Each load component has a unique frequency range. A load decomposition is made for the load sequence using digital filters with different passband frequencies. After load decomposition, each load component can be forecasted by an Adaline. Each Adaline has an input sequence, an output sequence, and a desired response-signal sequence. It also has a set of adjustable parameters called the weight vector. In load forecasting, the weight vector is designed to make the output sequence, the forecasted load, follow the actual load sequence; it also has a minimized least mean square error. This approach is useful in forecasting unit scheduling commitments. Mean absolute percentage errors of less than 3.4% are presented from five months of utility data, thus demonstrating the high degree of accuracy that can be obtained without dependence on weather forecasts.
A hybrid ANN-fuzzy expert system is developed [PS41] to forecast one hour to forty-eight hour ahead electric load accurately. The fuzzy membership values of load and other weather variables are the inputs to the ANN and the output comprises the membership value of the predicted load. An adaptive fuzzy correction scheme is used to forecast the final load by using a fuzzy rule base and fuzzy inference mechanism. The paper also presents a fuzzy pattern classification approach for identifying the day-type from the historical load database to be used for training the ANN. Extensive studies have been performed for all seasons, although the results for a typical winter day are given in the paper to demonstrate the powerfulness of this technique.
The paper [PS42] analyses the application of Kohonen's self-organizing feature map to short-term forecasting of daily electrical load. The aim of the paper is to study the feasibility of the Kohonen's self-organizing feature maps for the classification of electrical loads. The network not only 'learns' similarities of load patterns in a unsupervised manner, but it uses the information stored in the weight vectors of the Kohonen network to forecast the future load. The results are evaluated by using several months of hourly load data of a real system to train the network, and forecasting the daily loads for two periods of one month. The method is then improved by adding a second type of ANN for weather sensitive correction of the load previously calculated with the Kohonen network. This second type of network is a one-layered linear delta rule network.
The paper [PS43] addresses short-term load forecasting using machine learning and ANN techniques. ANNs, though accurate in weekday load forecasting, are poor at forecasting maximum daily load, weekend and holiday loads. This necessitates development of a robust forecasting technique to complement the ANNs for enhanced reliability of forecast and improved overall accuracy. The statistical decision tree method produces robust forecasts and human intelligible rules. These rules provide understanding of factors driving load demand. Decision trees when combined with ANN forecasts, produce robust and accurate forecasts. Simulations are performed on a service area susceptible to large and sudden changes in weather and load. Forecasts obtained by the proposed method are accurate under diverse conditions.
The paper [PS44] proposes a recurrent ANN based approach to short-term load forecasting in power systems. Recurrent ANNs in multilayer perceptrons have an advantage that the context layer is able to cope with historical data. As a result, it is expected that recurrent ANNs give better solutions than the conventional feedforward multilayer perceptrons in term of accuracy. Also, the differential equation form of the time series is utilized to deal with the nonstationarity of the daily load time series. Furthermore, this paper proposes the diffusion learning method for determining weights between units in a recurrent network. The method is capable of escaping from local minima with stochastic noise. A comparison is made between conventional multilayer perceptrons and the proposed method for actual data.
The paper [PS45] compares the ability of six ANNs to predict hourly system load for the Puget Sound Power and Light Company, a major North American electric utility. The neural nets, along with four other types of models, were used to forecast hourly system load for the next day on an hour by hour basis. This was done for the period November 1, 1991 to March 31, 1992.
The paper [PS46] discusses an ANN model for short-term load forecasting. A two-step training method to cope with a shortage of training data and overfitting problems is proposed. A limit is conducted to the range where the ANN's weights are allowed to change in order to preserve the general relation between the inputs and the output of the ANN. The ANN trained with this two-step training method demonstrates improved accuracy over conventional methods, including ANNs which employ ordinary training algorithms.
An approach to electric load forecasting which combines the powers of ANN and fuzzy logic techniques is proposed [PS47]. Expert knowledge represented by fuzzy rules is used for preprocessing input data fed to an ANN. For training the ANN for one-day ahead load forecasting, fuzzy if-then rules are used, in addition to historical load and weather data that are usually employed in conventional supervised learning methods. The fuzzy front-end processor maps both fuzzy and numerical input data to a fuzzy output. The input vector to the ANN consists of these membership values to linguistic properties. To deal with the linguistic values such as high, low, and medium, an architecture of ANN that can handle fuzzy input vectors is propounded. The proposed method effectively deals with trends and special events that occur annually. The fuzzy-ANN is trained on real data from a power system and evaluated for forecasting next-day load profiles based on forecast weather data and other parameters. Simulation results are presented to illustrate the performance and applicability of this approach. A comparison of results with other commonly used forecasting techniques establishes its superiority.
The paper [PS48] describes a solution of implementing neural capabilities into an energy management system. Starting at the point of necessity of increasing the forecast accuracy, a powerful combination of unsupervised and supervised learning networks for very short-term load forecasting is described as an example. The hardware and software components of the forecast system are also explained.
Send Questions, Comments, and Suggestions to bencr@hotmail.com