The Medium And Short-term Solar Irradiance Forecasting Research Based On Machine Learning Algorithms

With more and more large-scale photovoltaic(PV)incorporated into the power grid,the intermittency and volatility of PV have exerted negative impact on the safe and stable operation of the power grid.Solar irradiance is the most important factor to determine the fluctuation of PV.Therefore,accurate forecasting for solar irradiance is very important to PV,which will be conducive to real-time power dispatching and plan-ning of PV system to ensure safe and stable operation of power grid.This paper proposed three solar irradiance forecasting models based on machine learning algorithms and several real-world datasets.The specific implementations are as follows:(1)The traditional machine learning algorithms—support vector machine(SVM),multilayer perceptron(MLP)and recurrent neural network(RNN)for irradiance fore-casting are researched.These forecasting models are examined by using a real-world dataset.In experiment,the input vector is divided into two types:A)historical irradi-ance data;B)historical irradiance data and meteorological data.The experimental re-sults show that the forecasting performances of group A is better than that of group B but both groups have poor prediction performance.The root mean square error(RMSE)of group A models is between 89.71 W/m~2 and 94.34 W/m~2.The RMSE of group B models is between 86.41 W/m~2 and 87.23W/m~2.(2)A day-ahead irradiance forecasting model based on gated recurrent unit(GRU)and weather forecast data is proposed.The model takes the weather forecasting factors as the input vector,and then GRU network employed to forecast 24 h irradiance in the next day.The results show that GRU model reduces the RMSE of the CSIP,MLP,and RNN models by 16.58 W/m~2,37.14 W/m~2 and 48.50 W/m~2,respectively.Furthermore,compared with the long-short term memory(LSTM)model,the training time of the GRU model is reduced by nearly 40%.(3)A hybrid complete ensemble empirical mode decomposition adaptive noise(CEEMDAN)and deep learning network(CNN-LSTM)model is proposed.In this model,CEEMDAN decomposes the historical sequence of irradiance into relatively simple parts as the input of CNN-LSTM,and then the output is calculated by CNN-LSTM.The average RMSE of CEEMDAN-CNN-LSTM in regions of America and Algeria is 40.28 W/m~2.CEEMDAN-CNN-LSTM reduces the RMSE of other seven benchmark models by 11.60%-70.92%.