Time Series Short-term Forecasting Method And Its Application In Renewable Energy Field

Time series data widely exist in different fields,and its prediction has always been a hot research topic.The core of time series forecasting method is to extract the change law from the sequence and estimate the future data.With the development of data mining and machine learning methods,time series forecasting methods are becoming more and more advanced.However,each time series data have different changing laws,so the forecasting method is not universal.According to the characteristics of each time series data,apply the corresponding data processing and forecasting methods can improve the accuracy.Time series of wind speed and solar irradiance are typical data in the field of renewable energy,and are extremely important for wind and solar power generation.According to the characteristics of wind speed and solar irradiance data,the short-term forecasting methods are studied and discussed in this study.The main work and innovations are as follows:(1)Aiming at the large amount of wind speed time series data,two active learning strategies are proposed to select typical samples for short-term wind speed prediction.The proposed methods combine sample information and model information,and optimize training sample set by using active learning strategy.The experimental results show that the two proposed active learning algorithms can effectively optimize the samples,reduce the complexity of the model and ensure the accuracy of the wind speed forecasting model.(2)Aiming at the slow time-varying characteristics of wind speed data,a hybrid nonlinear estimation method combining Gaussian Process(GP)and Unscented Kalman filter(UKF)is proposed to handle the dynamic changes and improve the forecasting accuracy.The proposed method uses UKF to solve the nonlinear state space equation established by the GP model.This method can provide both point prediction and probability prediction of wind speed.In order to evaluate the performance of the proposed method,the wind speed data from three different wind farms are taken as the experiment object.The results show that the proposed method has improvement in point prediction and probability prediction.(3)Aiming at the problem that the direct normal irradiance(DNI)is affected by many factors,a short-term forecasting method of DNI based on real-time clear-sky model and cloud cover is proposed.The theoretical clear-sky irradiance is an important parameter for DNI prediction.Real-time adjustment of Linke coefficient is realized by identifying clear-sky points,and more accurate theoretical clear-sky irradiance is obtained.A short-term forecasting model is established by using historical DNI,cloud cover and theoretical clear-sky irradiance.Experiments data is provided by NREL open database.The results show that the proposed method can improve the forecasting accuracy of DNI.(4)How to extract effective features of the whole sky image is a difficult problem.To solve this problem,a short-term DNI forecasting model based on deep learning algorithm is proposed.Firstly,3D-CNN is used to process several continuous ground-based cloud(GBC)images and extract cloud texture and motion information automatically.Then,a DNI forecasting model is established by combining the cloud features,the theoretical clear-sky irrandiance and the past DNI.The experimental results show that the cloud feature extracted by 3D-CNN can reflect the attenuation of DNI well.By adding cloud feature as model input,the accuracy of forecasting model can be improved.(5)Aiming at the problem of multi-type features fusion in DNI forecasting model,fuzzy inference system(FIS)models based on hybrid fuzzification are proposed for short-term DNI prediction.Sky image features and numerical time series of irradiance are used as inputs of forecasting model to improve accuracy.The proposed FIS-tree and adaptive neuro-fuzzy inference system(ANFIS)model can overcome the problem of multi-type feature fusion in different dimensions by using hybrid fuzzification.The experimental results show that the hybrid fuzzification is more effective than the simple normalization,and the prediction accuracy is improved by the proposed FIS model based on multi-type feature fusion.