CNN-based LSTM photovoltaic power generation power prediction model and construction method thereof

Abstract

The invention belongs to the field of photovoltaic power prediction, and discloses a CNN-based LSTM photovoltaic power generation power prediction model and a construction method. The model onsists ofa two-layer CNN convolutional layer, an LSTM, an attention mechanism and a full connection layer. When only power data exists, feature extraction is carried out on an input one-dimensional power generation power time sequence through a double-layer convolution layer, a plurality of data matrixes are obtained through 3 * 1 convolution kernel operation. Information abstraction from the time sequence to high-level features is achieved. The attention mechanism automatically pays attention to an LSTM hidden layer output vector obtained by each matrix through an LSTM algorithm, and a greater weightis given to a characteristic quantity remarkably related to the current output quantity. The output vector of the attention mechanism is processed into a one-dimensional vector through an unfolding operation and inputted the one-dimensional vector into the full connection layer, wherein a full connection layer directly outputs the predicted value of the photovoltaic power generation power at thenext moment. The method is small in data size, high in transportability and high in prediction precision.

Description

technical field [0001] The invention belongs to the technical field of forecasting or optimization, and in particular relates to a CNNLSTM-based photovoltaic power forecasting model and a construction method thereof. Background technique [0002] Currently, the closest prior art: [0003] The common photovoltaic power prediction methods mainly include physical method, regression method, time series method and machine learning method. The physical method uses geographic information, photovoltaic module parameters, sunlight irradiance, atmospheric temperature and other parameters to construct a physical prediction model, and the physical method relies on the meteorological parameters of numerical weather prediction (NWP) to predict the power of photovoltaic power generation. Japanese scholars proposed a physical prediction model that uses solar radiation intensity as an input to predict the electrical energy output of photovoltaic arrays; the simple physical prediction model ...

AI Technical Summary

Problems solved by technology

As a typical traditional machine learning algorithm, the BP neural network is an unstable model, which is easy to fall into local optimum and slow iterative convergence, and the setting of the number of neuron layers in the BP neural network is easy to cause overfitting and degradation phenomena

SVM is a stable model that is not easy to fall into local optimum, but SVM is only suitable for dealing with small sample problems, and it also has the problem of slow iterative convergence

Traditional machine learning algorithms such as BP neural network and SVM rely on feature engineering, and the quality of sample features directly determines the prediction accuracy of the model. At present, artificially designed features are still in a dominant position, and artificially designed features need to rely on the prior knowledge of the designer. Manually adjust parameters, so only a small number of parameters are allowed as sample features. Usually, the monitoring data of photovoltaic power plants or NWP data are directly used as feature parameters. Traditional machine learning algorithms cannot effectively extract deeper information from these feature parameters, and the accuracy of the prediction model is still improved. It is impossible to predict the future photovoltaic power generation capacity relatively accurately, which will lead to the impact on the stability and security of the power grid when large-scale photovoltaic power is integrated into the grid. Photovoltaic power generation prediction is the key to eliminating threats and maintaining the safety and stability of the power grid one of the techniques

Traditional machine learning causes low prediction accuracy. Take BP neural network as an example. In the field of photovoltaic power generation with a huge amount of data, BP neural network can only construct a shallow neural network. If the number of neuron layers is too large, it will affect the accuracy. As a result, the features in the data are not fully utilized, and the overall accuracy is not high compared with deep learning.

Taking SVM support vector machine as an example, SVM is suitable for small-scale data analysis. When the amount of data is huge, the convergence speed is slow, and it is difficult to apply it in the field of photovoltaic power generation prediction.

[0019] To sum up, the problems existing in the existing technology are: traditional machine learning algorithms rely on manually designed sample features, and there are shortcomings in feature extraction, resulting in low prediction model accuracy

Attention mechanism Attention mechanism has been widely used in image processing, video analysis and natural language processing, but less in the field of photovoltaic power generation prediction

Images