Research Of Office Building Power Prediction Based On The Improved RBF Neural Network

Facing the current situation of the fast growth of public building energy consumption and the irrational use of energy,realizing accurate and rapid prediction is helpful for managers to find outliers in time and develop economical and reasonable energy-saving measures.In recent years,the neural network methods have been widely applied to the construction of building energy consumption prediction model.In this paper,the shortcomings of the radial basis function neural network were improved to establish new prediction models which was applied to the power consumption prediction of office buildings.The main contents are as follows:For the uncertainty of k-means algorithm in determine the center of RBF neural network,the AP-RBF neural network model was proposed to forecast the power consumption of office building,which introduced affinity propagation clustering algorithm to obtain the high quality clustering center as the center vector of the hidden layer,then the neural network parameters were optimized by gradient descent method.The experimental results show that,compared with BP neural network and k-means RBF neural network model,new model improves the stability greatly.In order to further improve the effectiveness of the prediction model,the particle swarm optimization algorithm was used as learning algorithm to optimize the node center,width and connection weights of RBF neural network.The neural network prediction models were built based on the LDIWPSO algorithm,LIIWPSO algorithm and FWPSO algorithm respectively.The experimental results show that,the LDIWPSO-RBF forecast model is more stable and accurate than LIIWPSO-RBF,FWPSO-RBF and k-means RBF neural network model.Considering the particle swarm optimization which may fail into local extremum in the course of learning and training,the particle's fitness value was introduced into the inertia weight formula.By comparing the difference between the individual and the overall average level,the particles were adjust the step size dynamically and adaptively,then jumped out of the local optimum.Meanwhile,aiming at the shortage of the complexity of particle dimension,the AP clustering algorithm was added to determine the number and location of the center which form APFPSO algorithm.Four test functions were processed to verify the progressiveness of the improved algorithm.Then the new RBF neural network prediction model based on APFPSO was built.Finally,the power consumption of office buildings in the next 30 days were predicted,the results show that the APFPSO-RBF model contribute the accuracy rate more than 92%,and overcome the shortcomings of the RBF neural network and PSO algorithm.