Research On Short-term Electricity Load Forecasting Based On Neural Network

Electricity load forecasting result has been the critical working basis of planning, dispatching and marketing departments in power system for a long time. Guaranteeing that the forecasting is scientific, reliable and precise is of great importance to the correct decisionmaking of power system.In this paper, four different models for load forecasting, which use different theories, are built on the basis of a thorough analysis of the characteristics of power load. The models are built respectively, based on BP neural network, Elman neural network, RBF neural network and wavelet neural network. The four models are applied in an example of a city in HuBei province for validation and the results are compared for analysis. The results show that all of the four models can reach some certain precision in forecasting, while great improvement can still be made.For the purpose of improving the precision of forecasting, this paper comes up with a hybrid training algorithm for neural network which combines particle swarm optimization, genetic algorithm and improved ant colony algorithm together. The algorithm mainly focuses on how to diminish the disadvantages of a new type of neural network using ant colony algorithm for training and makes three kinds of improvement. Firstly, aiming at solving the problem that taking values in subinterval of weights is of no basis and thus values are randomly taken, the concept of candidate groups is introduced. And particle swarm optimization is introduced into the algorithm to optimize the values in weighs candidate groups, which reduces the error and increases convergence rate of the algorithm. Secondly, to solve the problem of low rate in the beginning of the search, genetic algorithm, which has remarkable ability of global optimization, is introduced to help determine the initial distribution of pheromone. This improvement greatly reduces the time taken in initial stage of optimization and keeps the algorithm from being trapped in local optimum. Thirdly, the updating strategy of pheromone used in fundamental ant colony algorithm is also changed. Apart from using elite ant, rank-based ant system and constraints of pheromone concentration, non-elite ants are also allowed to release pheromone. This little change can greatly reduce the possibility of premature convergence of ant colony algorithm. The above three improvement strategies improve the performance of ant colony algorithm greatly, compared with the fundamental ant colony algorithm.Based on this hybrid algorithm, a short-term forecasting model is built and it goes through a practical calculation example to test its validity. After comparing the forecasting results of this improved hybrid forecasting model with the results of typical neural network forecasting methods and neural network model using fundamental ant colony algorithm, the result shows that, the neural network model using improved hybrid ant colony algorithm has the best performance and the precision measures up to the standard formulated by State Grid Corporation.