Improved Echo State Network And Its Application In Solar Energy Prediction

Solar hybrid power generation systems play an important role in the utilization of new energy and saving electricity costs.Accurate prediction of solar energy can provide a guarantee for the safe and stable operation of those systems,which is of great significance to save power costs and improve economic benefits.Due to the advantages of good dynamic characteristics and simple and efficient modeling efficiency,echo state network has become a novel research hotspot in the field of time series prediction.However,the traditional echo state network is a single-reservoir single-task learning model composed of an input layer,a hidden layer(reservoir)and an output layer.It has limited computational capability,poor scalability and robustness,and is not suitable for those complex solar energy prediction tasks with multiple timescales and multi-task needs,which limits its wide application and development in practical engineering to some extent.Therefore,the solar energy time series is taken as the research object,and echo state network is chosen as the main methodology in this paper.From the structural design and improvement,different echo state network algorithms are developed to predict solar energy in different task needs.The main contents and innovations of this paper are as follows:(1)A multi-clustered echo state network algorithm is designed to predict single-step and multi-step solar energy.Based on the complex network theory,a multi-cluster reservoir topology is firstly designed.In order to explore the influence of different historical information and temperature and humidity factors on solar energy modeling accuracy,different input-output multi-cluster echo state network prediction models are constructed,respectively.In addition,most of the existing studies only make quantitative analysis on the predicted results in experimental simulation,ignoring some internal characteristics that can reflect the law of data evolution.Therefore,the validity of this established multi-cluster echo state network is verified by the combination of quantitative and qualitative analysis.Quantitative analysis includes root mean square error,mean absolute error and so forth,which can externally reflect the accuracy of predicted results.Qualitative analysis includes non-stationarity,complexity and so forth,which can internally reflect the accuracy of predicted results.(2)A multi-timescale echo state network algorithm is proposed to fulfill multi-task solar energy prediction needs.Aiming at the solar energy prediction problem with muti-timescale and multi-task requirements,the multi-timescale prediction framework is proposed.Under this framework,a multi-timescale echo state network algorithm is designed to meet multi-task solar energy prediction needs.Based on the historical solar energy information and multi-timescale echo state network,a new type of solar energy model is established that can fulfill multiple prediction tasks in parallel.In this model,related information on different time scales is shared through multiple reservoirs,and the expected value is integrated in the output layer.Compared with single-timescale single-task prediction models,the proposed model can fulfill multiple tasks in parallel more effectively.Furthermore,the proposed multi-task model is also superior to the single-task model in terms of modeling accuracy,because the related information between tasks on different timescales are shared by each other.(3)A multi-reservoir echo state network algorithm under the concept of deep learning is developed to meet the simple and efficient demand of solar energy prediction.The classical single-reservoir echo state network has limited computational ability,and the existing deep learning models have problems such as heavy computational burdens.Therefore,in order to take advantage of the powerful learning ability of deep learning and the simple and efficient training mechanism of echo state network,a multi-reservoir echo state network algorithm is developed for solar energy prediction.Based on the historical solar energy information,a multi-reservoir echo state network model is established for single-step and multi-step solar energy prediction.The influence of hidden layers on the network performance is analyzed,when the total size of reservoir is constant.From the internal features,the reasons for network performance differences under different hidden layers are further discussed.Simulation results demonstrate that the designed multi-reservoir echo state network performs better than single-reservoir echo state network under constant reservoir size,with richer dynamic characteristics and stronger information processing capability.(4)A chain-structure echo state network algorithm with better robustness is proposed to conducted spatio-temporal prediction of solar energy.The classical echo state network has limited ability to integrate the features of large high-dimensional data sets.In addition,the scalability and robustness of one model also need to be further studied,when the dimension of input variables changes or the signal from a certain dimension suffer from disturbance or interruption.Therefore,a novel chain-structure echo state network algorithm is proposed in this section,in order to improve the scalability and robustness of traditional echo state network in spatio-temporal modeling.Based on the historical spatio-temporal solar energy information,the chain-structure echo state network model is developed for single-step spatio-temporal prediction of solar energy.Motivated by the philosophy of `divide and conquer`,the proposed model is composed of multiple ESN modules,among which each echo state network module is responsible for mapping dynamic feature of each cluster of input vectors.By successively conquering the predicted values of each clustered variable,the final predicted results are then obtained.Compared with traditional echo state network and other benchmark models,experimental results show that the proposed model has better modeling accuracy,scalability and robustness to deal with high-dimensional variables.