Research On Photovoltaic Power Prediction Based On Attention Mechanism And Multi-task Learning

Solar energy is one of the clean energy sources,but in the process of constructing micro-grid or photovoltaic power generation and grid operation,its volatility and randomness will affect the operation of the grid,so it is necessary to accurately predict the photovoltaic power generation.The application of deep learning algorithms in the field of photovoltaic power generation prediction has been studied for many years,but there is still room for improvement in prediction accuracy and response speed.This paper studies the application of deep learning algorithms in the field of photovoltaic power generation prediction,and proposes a prediction algorithm based on the combination of self-attention mechanism and multi-task learning,aiming to construct a model which has fast convergence speed,high generalization ability,and have a good balance in prediction accuracy and response time.The main innovations of this article are:(1)A method of applying the self-attention mechanism to the field of photovoltaic power generation forecasting is proposed.The important information of sparse data is extracted by calculating the similarity between features.Compared with the traditional time series model RNN needs to rely on the order between features,the method can better dig out the relationship between features and improve the prediction accuracy of the model.At the same time,the self-attention mechanism captures the internal correlation of features through matrix operations.This computing feature can achieve parallelism calculation,which improves the calculation efficiency of the model.(2)It reveals the feature extraction function of the Transformer-based self-encoding network applied to the field of photovoltaic power generation prediction.The encoder receives the input and compresses it into an internal representation,and the decoder reconstructs and outputs the internal representation.By optimizing the internal representation can reduce the gap between the output and the original input,and achieves the purpose of promoting the model to extract more effective features.(3)A multi-task learning model is constructed.One of the tasks is the task of extracting high-efficiency features from the self-encoding network.The second task is to add a prediction module to predict photovoltaic power generation based on the extraction of high-efficiency features from the self-coding network.The multi-task learning model speeds up the convergence speed and improves the accuracy of model prediction through mutual cooperation and sharing of parameter mechanisms.Finally,the algorithm in this paper not only has good predictive performance,but also is suitable for resource-constrained platforms.Compared with the photovoltaic power generation prediction model that uses CNN and LSTM as the feature extractor,the error of the model established in this paper is reduced by 14.82%and 8.09%respectively,which has a wider range of application value.