Mid-term Prediction Of Photovoltaic Power Based On Long And Short-term Memory Network

As traditional fossil energy sources are facing increasingly serious reserves crisis and environmental pollution problems,people have shifted the focus of future energy development to clean energy represented by wind energy and solar energy.As a very important utilization method of clean energy,photovoltaic power generation is widely developed and utilized because of its environmental friendliness and no need for any "fuel".However,under the influence of external factors such as irradiance and temperature,the output of photovoltaic power plants shows great fluctuations and randomness,which affects the safe stable operation of power grid seriously.Therefore,it is necessary to carry out accurate photovoltaic power prediction.The current forecast time scale for solar and wind energy in my country is 3 days,which does not match the start-stop cycle of thermal power from 5 to 7 days.The grid dispatching department cannot incorporate new energy and thermal power into the start-up plan for guidance.In response to this problem,this paper proposes a medium-term photovoltaic power prediction method based on long-and short-term memory networks,which extends the time for photovoltaic power prediction to the next seven days,in order to reduce the impact of photovoltaics on the grid and promote the consumption of photovoltaics.The main research contents are:1.Establish LSTM photovoltaic power prediction model.The selection of LSTM neural network avoids the shortcomings of the traditional statistical algorithm’s insufficient generalization ability and easy to fall into the local optimal problem,and the deep learning algorithm RNN network often faces the problem of gradient disappearance and gradient explosion in the process of long-term sequence learning.Through preliminary analysis of the corresponding relationship between various meteorological factors and photovoltaic output power curve,the application of mutual information theory and Pearson correlation coefficient algorithm to screen the characteristics of meteorological factors,so as to determine the irradiance,temperature and humidity as the input of the prediction mode.Input to the model.After setting the number of input and output layers of the LSTM neural network,combining empirical formulas and trial and error methods to determine the number of hidden layers,the LSTM neural network model is constructed.2.The cloud model theory is applied to the selection of similar days,and the cloud model-LSTM photovoltaic power prediction model is constructed.Firstly,the daily irradiance data is filtered from the historical data,and then the irradiance cloud model is constructed.Use cloud transformation and expectation curve method to calculate and select similar days to the forecast day.The verification was carried out in four seasons,and the results show that the method can select dates that are similar to the predicted daily average irradiance and the overall change trend of irradiance.3.Considering seasonal differences,randomly select seven days in spring,summer,autumn and winter,and predict and compare the constructed cloud model-LSTM photovoltaic power prediction model together with a single LSTM,SVM,and GM model.The results of using MAPE and RMSE indicators to quantify the prediction error show that the cloud model-LSTM photovoltaic power mid-term prediction model constructed in this paper has good accuracy.