Research On Photovoltaic Power Forecasting Based On Machine Learning

Photovoltaic power forecast can be used in power grid scheduling,fault detection,etc.It is one of the key technologies for grid-connected photovoltaic power plants.In this paper,the research of photovoltaic power forecasting model is the main line,and the popular machine learning technology is applied to the modeling and forecasting of photovoltaic power data.The main work and research are as follows:(1)Five popular PV power forecasting methods were analyzed,including physical method,regression method,grey forecasting method,time series forecasting method and machine learning method.The characteristics,advantages and disadvantages of these methods were analyzed.The combination of Kmeans and SVM was researched for short-term photovoltaic power forecasting.The Kmeans algorithm was used to cluster training samples to reduce noise interference caused by other types of samples.Each type of sample trains a SVM for photovoltaic power forecast.The forecasting effect of the kmeans-SVM model on sunny,cloudy and rainy days was investigated.The experiments show that the Kmeans-SVM model can effectively forecast the photovoltaic power under three meteorological conditions,and the forecast performance is better than comparison models,Kmeans algorithm can effectively improve the forecast accuracy.(3)The method of ensemble learning were introduced to research the short-term photovoltaic power forecast by integrating multiple SVMs with Stacking algorithm.The experimental results show that the Stacking-SVM model has outstanding forecast performance under cloudy conditions,and the forecast accuracy under sunny and rainy conditions is also significantly improved.Ensemble learning can significantly improve the forecast performance of single learner model.(4)The ALSTM model which combines LSTM algorithm with attention mechanism was used to forecast photovoltaic power.This model uses the photovoltaic power time series and the photovoltaic module temperature time series as input to forecast the photovoltaic power at the next moment.ALSTM uses two LSTMs to extract the features of the two time series respectively,and uses the attention mechanism to pay attention to the useful items in the LSTM hidden layer output vector.Experiments were carried out using one-year measured photovoltaic power data.The forecast result of ALSTM model in spring,summer,autumn,winter and different time horizons were investigated.The experimental result show that ALSTM model can make effective forecast within 60 minutes.The attention mechanism can significantly improve the the quality of feature extraction and forecast accuracy.(5)The combination of machine learning and big data technology to forecast PV power was researched.The Spark on YARN experimental platform was built,and the decision tree regression,GBDT regression and random forest regression provided by Spark Mllib were used to model the photovoltaic power data,and experiments show that random forests get better forecast results.