Ultra-short-term Photovoltaic Power Prediction Based On TCN-LSTM Model

To promote energy transformation and upgrading,help sustainable development and realize the double carbon strategic goal,our country’s new energy development is speeding up year by year.Among them,photovoltaic power generation plays an important role in new energy power generation due to its wide distribution of energy sources,and its development speed is particularly rapid.But at the same time,due to the strong volatility of photovoltaic power generation,a large number of photovoltaic power stations connected to the grid will pose a certain threat to the stable operation of the power system.Therefore,the ultra short term prediction of photovoltaic power plays an extremely important role and value in the safe and stable operation of power system and the healthy development of photovoltaic industry.In order to achieve ultra-short term prediction of photovoltaic power with higher precision,this paper mainly studies the following aspects on the basis of extensive reference to relevant data and literature:First of all,abnormal data and low-value information often exist in the original data related to photovoltaic power,but the original data is a very important basis for photovoltaic power prediction.For the preprocessing of original data,this paper analyzes the influence of meteorological factors on photovoltaic power based on the timing curve generated by various data and Pearson correlation coefficient.Then,the outliers of the original data were detected from the two aspects of data distribution characteristics and correlation characteristics,and the boxplot method and 3σ method were used to detect the outliers of each data on the time series distribution.Based on correlation analysis,the experimental comparison of Support Vector Machines(SVM),Isolation Forest(i Forest)and Density-Based Spatial Clustering of Applications with Noise(DBSCAN)for the detection effect of photovoltaic power and total irradiance outliers,the i Forest with the best detection effect was selected as the further detection method for outliers.In the process of replacement optimization of outliers,the optimization effects of mean filling,median filling and upper and lower mean filling were compared through experiments.According to the evaluation index results,the upper and lower mean values were selected as the replacement filling method of outliers.Based on the data information after outlier processing and Pearson correlation coefficient,meteorological data and power data with high value are selected as inputs to effectively complete the optimization processing of the original data.Secondly,in order to select a good neural network as the basis of prediction model,this paper is based on Back Propagation(BP)neural network,Elman neural network,Long Short Term Memory(LSTM)neural network,three kinds of photovoltaic power prediction models are established respectively.With the pre-processed data information as input,the comparison experiment of photovoltaic power prediction is carried out.Through the analysis of the experimental results,the LSTM prediction model has better prediction accuracy than the BP prediction model and the Elman prediction model.Finally,considering the long sequence of photovoltaic related data and the difficulty in extracting feature information from the model,this paper adds the Temporal Convolutional Network(TCN)into the LSTM prediction model by means of combinatorial optimization,and establishes the TCN-LSTM photovoltaic power prediction model.The feature extraction ability of the input long time series information is improved.The experimental results show that the TCN-LSTM prediction model has better prediction accuracy than LSTM prediction model,BP prediction model and Elman prediction model.Since the complexity and non-stationality of photovoltaic power sequence will bring some difficulties to the feature information processing of the prediction model,this paper has added Variational Mode Decomposition(VMD)into the TCN-LSTM photovoltaic power prediction model.The photovoltaic power prediction model of VMD-TCN-LSTM was established,and the photovoltaic power sequence was decomposed into trend component and other components by VMD,which effectively reduced the complexity and non-stationary of photovoltaic power sequence.Through experimental verification,VMD-TCN-LSTM prediction model has higher prediction accuracy than TCN-LSTM prediction model,BP prediction model,Elman prediction model and LSTM prediction model,and better realizes ultra-short-term prediction of photovoltaic power.Through the experimental analysis of four kinds of seasonal data,it is verified that the VMD-TCN-LSTM prediction model has good adaptability and stability.