Distribution Network Short-term Load Forecasting Based On VPSO-ELman Neural Network

The accuracy of the distribution network's Short Term Load Forecasting with a direct impact on system distribution network system planning, operation scheduling, production planning and quality of power supply. Gradually, the power supply management will realize maeketization, Short-term power load forecasting has become a important symbol of modernization to measure a Power supply enterprises. So, seeking to achieve an optimal accuracy of a load forecasting method has important research significance and practical value.Because the load date's acquisition rely on SCADA which may be failure can result in anomaly dates, and other unusual factors can brought about brought about too. In this paper, using the method of mathematical statistics exception identification, and using three-point smoothing method to amend these unusual dates. The given practical example proved that this method is simple to program, fast and has small work load. Through the algorithm processing the data in the sample after the training, greatly improving the load forecast accuracy.Artificial Neural Networks(ANN) as an intelligent way, in recent years, has been widely used in power system load forecasting, and achieved good results. However, in a large number of applications, mainly using the BP neural network, a static neural network for predicting the dynamic load process, easy to fall into local minimum, the prediction accuracy is difficult to have substantially enhanced. In this paper, using Elman network which is a dynamic recurrent neural network model and use BP algorithm. For making the network's work better, proposed adaptive learning rate and momentum method with the BP algorithm. Application of improved BP algorithm can not only reflect the short-term load's nonlinear characteristics, but also reflect the dynamic process, and to a certain extent, improve the convergence rate of the ELman in the training process and avoid falling into local minima. Use a city's distribution network's load date which belong to Shandong power grid, by studying and analyzing these date, and apply these data to train Elman network and predict the load value 12 times points of the day. by comparison with BP network, proved ELman model is superior to BP model.Because the Elman network use BP algorithm too, In large-scale short load forecasting process, Elman network will still inevitably some inherent flaws which BP network has. In this paper, using the Variance Particle Swarm Optimization to optimize and train the ELman network. Based on Particle Swarm Optimization model, when the particle's speed below a certain threshold, VPSO algorithm will provide a appropriate impulse to particle which can re-initialization the particle and re-position the particle's location. This would avoid the PSO algorithm into a local optimum of the defects, to achieve the global optimum. Application VPSO algorithm optimized Elman neural network, to use the weights and thresholds between each layer of the Elman network connection as the VPSO's particles. To optimize the initial distribution of these threshold and weights, when the optimization to a better location, in the local optimization using neural network. The optimized Elman network can avoid the convergence speed and getting into local minimum solution and other defects to the full. Though using MATLAB to regional power grid's load forecasting, This case confirms the effectiveness of the proposed method, obtain a higher prediction accuracy, the Mean Average absolute Error are 1.073%, and the maximum relative error are 2.12%, Predictive performance of VPSO is better than single ELman network's. So, this method is fully meets the requirements of the load forecast.