Kalman Filter Learning Algorithm For Feedforward Neural Network And Its Application In Short-term Load Forecasting

The classical method for training a multilayer feed-forward artificial neural network is the back-propagation(BP) algorithm. Although it is successfully used in many cases and has been improved continuously, the BP algorithm suffers from a number of shortcomings. One is the slow convergent rate. Another shortcoming is the convergence may be local. These shortcomings have not been got over radically. In this thesis a learning algorithm based on Kalman filter is proposed for training a neural network. Kalman filter is an optimal estimate method. It describes the filter using a state space model of a linear stochastic system which is made of state equation and measurement equation and estimates the state variable optimally according to the rule of linear agonic and minimum root mean squared covariance.Kalman filter training algorithm is to estimate the network weights as the state vectors of the Kalman filter. Because the extended Kalman filter training algorithm estimates the weights according to the rule of minimum root mean squared covariance, it need less iteration than BP algorithm. Further more, not involving convergent parameters made it easier to apply.In the extended Kalman filter, the covariance matrix will lost its positive definition due to the error rounding when calculating in computer. This will consequently lead to the instability of the numerical value and radiation of the filter. Singular value decomposition is applied in the calculation of the covariance matrix. This method improved the calculation rate and the stability of the numerical value.Artificial neural network is popular in short-term load forecasting. The main advantage of using neural network lies in its good nonlinear mapping capabilities between multi inputs and outputs, the adaptive leaning abilities and the distributed storing weights. These strong abilities made it possible for prediction taking into account numerous weather conditions such as temperature, humidity, rainy and blizzard. These made the forecasting value more accurate than other methods. An alternative training algorithm of neural network based on Extended Kalman filter is used in application of the short-term load forecasting of Wuhan power system. The case proves that he algorithm proposed this thesis is effective.