An Improved Wavelet-Kalman Filter Based Short Term Load Forecasting

Some disadvantages of the existing Wavelet-Kalman filter based load forecasting methods are pointed out. An improved scheme is presented: A load is expressed as the product of an average component and a varying component. Such two components are forecasted separately. A new method based on artificial neural networks is proposed and applied in forecasting the average load. The average load is expressed as the summation of a temperature-insensitive part being stationary and a temperature-sensitive part. The former is obtained according to the data of temperature insensitive seasons. The latter is obtained by a BP network whose inputs are history information of load, temperature (including the averaged temperature, the highest temperature and the lowest temperature) and the temperature of the predict day. The varying component is forecasted by Wavelet-Kalman filter based method. The wavelet coefficients of the varying load component are obtained by the decomposition scheme of multi-resolution analysis. The wavelet coefficients are modeled as the state variables of the Kalman filter. The best estimation of the predicted varying load component is obtained by the recursive Kalman filter algorithm. Load forecasting results show that the improved method increase the forecast precision. A new method of real time modification on forecasting results is proposed. In case of several forecasted errors are over the threshold successively, the revise algorithm such as grey-forecasting is taken to modify the following forecasted data until a certain number of forecasted errors restore to the allowed range. Load forecasting results show that the revise method is satisfied. In addition, a bad data elimination method based on the wavelet analysis is adopted. By using the wavelet transform, the different load sequence components are projected to the different scales in which the matching modulus maxima can be obtained and eliminated according to the daily-period feature of the power system load. The effectiveness of the method is verified by the results of a practical example. An integrated short-term load forecasting program is completed based on above methods. The loads of Xi'an distribution network are used to test the program. Comparison of the results with the measured values show that the proposed methods and the program are feasible.