Short-term Heating Load Forecasting Based On RBF Neural Network

The problems of energy shortage and environmental pollution are becomingworse and worse. Central heating system becomes more and more popular becauseof reducing energy consumption, relieving environmental pollution and improvingprofits. With the revolution of central heating system, household metering heatingsystem has put forward high requirements for effective utilization of heating sourceand keeping pipeline balance. Hence, heating load forecasting is significant forscheduling the production of the heat-supply system.Firstly, the background and meaning of heating load forecasting are introduced,and the existing forecasting methods are analyzed and compared systematically.According to the long time lag and serious nonlinear characteristics of heatingsystem, the method based on radical basis function neural networks which hasstrong nonlinear mapping and generalization ability, is more suitable for heatingload forecasting.Secondly, in order to ensure the reliability of data, heating load datapreprocessing is necessary including elimination of abnormal data, improvement ofthe missing data, data normalization and de-noising processing. However, thetraditional methods deal with the data in one-dimensional space and have limitationsto some extent. Given horizontal and vertical continuity of load simultaneously, anovel method of identifying abnormal data in two-dimension space based on datadensity estimation is presented. The wavelet threshold de-noising method is usedand the de-noised data are demonstrated to improve the identification rate of data.Finally, a forecasting model based on radical basis function neural network isestablished after analyzing characters and inherent rules of heating load. In order tooptimize the model, an improved particle swarm optimization algorithm isintroduced to increase the convergence speed and avoid local minimum. Simulationresults show that it is feasible and practical by applying neutral network to forecastthe short-term heating load. The parameter and structure of radial basis functionneutral network are further improved by modified particle swarm optimizationalgorithm. The higher forecasting accuracy, output stability and the fast convergencespeed are achieved. In addition, the average relative error of the improved model isonly2.55%, which satisfies the requirement of accuracy rate within3%.