The Application Of Chaos Theory And Neural Networks In Short-Term Electric Load Forecasting

With the fast development of electric energy markets, electric load forecasting has become more and more important to the power industry . Accurate and immediate electric load forecasting is used to decide weather the output of an already running generation unit should be decreased or switched off. Similarly, forecasts are used to decide whether unnecessary rotating reserve capacity should be decreased. Meanwhile precision of forecast is also a basic requirement in keeping good stability in the running-process of power systems. In addition, forecasts are also crucial to make a good generators-maintaining plan. The last but not the least, forecasts of load have great impacts on keeping normal production and life, effectively reducing the cost of generation and improving the economic and social benefits.This thesis is structured as following:On the base of analysis to the real power system, the first chapter sums up not only the characteristics of real loads in power system but also the characteristics and types of electric load forecasting. Then some of classical and modern methods on electric load forecasting are discussed in the second chapter.And simple introductions for these methods are made. Moreover, the merits and defects of these methods also are analysed. Besides, some of the methods are applied to forecast electric loads.The third chapter proposes a modified chaotic forecast method.The chaotic forecast method is a method which is widely applied to electric load forecast.The use of chaos theory for electric load has the following steps. First of all, a phase space is reconstructed for historical data. Then a linear method is used to approximate the function in chaotic systems. Finally, the least square method is used to estimate the parameters of objective function. For the objective funtion is related with the similarities between the selected vectors and the base vector, the objective function has to be weighted in accordance with the degree of similarity with base vector. At present it generally uses the correlation degree to measure the similarity of vetors in the phase space. In order to reduce the computing time and computational complexity, a new method which uses vector 1th norm to measure the similarity of vectors in phase space is proposed. Then this method is used to forecast daily electric load. From the results we obtained, we can see that this method not only keeps a good precision but also reduces the computing time and computational complexity.In the fourth chapter RBF-AR model is also applied to electric load forecasting, which uses RBF to approximate the funtional coefficients in AR model. In this thesis SNPOM is applied to identify and optimize the parameters in this model.SNPOM separates the parameters into both parts: nonlinear and linear parts.The LMM method is used to optimize the centers.Similarly the LSM method is used to optimize linear weights.And in searching-process the structure of the parameter space is divided which is equal to reduce the parameter space.This arithmetic improve the convergence and precision. Finally, the RBF-AR model was applied to electric load forecasting. Then the forecasting results are given and compared with the previous methods. In the fifth chapter, the optimal combination of two methods improves the previous forecasting results, which is compared with those of other forecasting methods.