Construction And Application Of A Hybrid Forecasting Model Based On BP Optimized By Firefly Algorithm

With the excessive use of non-renewable energy sources,such as coal and oil,the living environment of human beings has been greatly damaged.Because of its recyclable,pollution-free,clean,wind has drawn widespread attention.As one of the most important indicators of wind energy,accurate wind speed forecast will bring great help to wind power production management and its dispatching.Therefore,the research on wind speed prediction method has drawn much attention.In recent years,BP neural network plays an important role in wind speed prediction.BP neural network's strong ability of nonlinear mapping,high self-learning and self-adaptability,as well as the ability of certain generalization and fault-tolerant ability have been favored by researchers.In particular,the strong ability of nonlinear mapping of BP neural network has high predictive performance in wind speed prediction.However,due to its shortcomings of local optimum,slow convergence and network training "over-fitting1",the prediction results of BP neural network are not stable.Therefore,its prediction effectiveness has great room for improvement.Researchers use ant colony optimization algorithm,genetic optimization algorithm,particle swarm optimization algorithm,firefly optimization algorithm etc.to optimize the BP neural network's weights and thresholds so as to improve its prediction performance.Among of them,the firefly algorithm is attracted by the researchers of all ages because of its global optimization and fast convergence.After optimizing,it not only improves the predictive performance of BP and enhances the predictive stability of the model.However,fireflies are apt to fall into the local optimum and have a slow convergence rate in the latter part of BP optimization,which has an impact on the prediction performance of the model.Obviously,desorption preprocessing of the original data before the prediction can improve the predictive performance of the model.Therefore,the data preprocessing can help promote the effectiveness of the model.In order to solve the above problems,this paper proposes a hybrid model on the basis of singular spectrum analysis,modified firefly optimization algorithm and back propagation(BP)neural network,which can better reduce the forecasting error and improve the prediction accuracy.To test the forecasting effectiveness of the proposed model,10-min wind speed collected from the 27th,28th and 29th wind turbine in Penglai City,Shandong Province,China is chosen to forecast the wind speed.The mean absolute percentage error(MAPE)is regarded as the main evaluation criterion to evaluate the predictive performance of the proposed hybrid model,and the other evaluation indexes are chosen as the reference.The smaller the MAPE value is,the higher the forecast effectiveness of the hybrid model is.The main contents of the proposed hybrid model are summarized as follows:1)The noise presented in the original data sequence is removed by singular spectrum analysis(SSA).The principle of this step is to decompose the original data and then remove the high frequency sequence of the original data,and the remaining contains which contain the important information of the original data are reconstructed to forecast the wind speed.2)Due to the non-linearity and irregularity of wind speed,the traditional forecasting method often possesses a higher forecasting error.And the neural networks have great advantage in forecasting the non-linear data,therefore,in this paper,we select the BP neural network which has been proved by practice and most widely applied to forecast the wind speed.3)As we all know,the single BP neural network is easily caught in local optimal while forecasting the wind speed,so we apply the firefly algorithm to optimize the weights and thresholds of the BP neural network so as to enhance its global optimization ability.4)For the firefly algorithm,individual firefly will be easily caught in local optimal and the FA will have a lower convergence rate while optimizing,which always leads to bad predictive results.Therefore,we apply the quasi-Newton method(BFGS)to modify FA,and the modified firefly algorithm not only retains the advantages of the firefly algorithm,but also overcomes the drawbacks while optimizing in the later stage.In this paper,we can conclude from comparison between the hybrid model and the other three single models as follows:a)The hybrid model has higher forecasting effectiveness than the single model,and the MAPE of the hybrid model are the smallest among all experiments,thus,we can learn that the hybrid model has a higher stability and reliability,b)The MAPE decreases substantially when SSA was used to remove the noise of the original time series data,which also confirms SSA can promote the forecasting accuracy.c)The results of the sub-season prediction,multi-step prediction,and DM(Diebold-Mariano)test further validate the hybrid model outperform the other single models.