Weighted fully-connected regression networks for one-day-ahead hourly photovoltaic power forecasting

Accurate photovoltaic power forecasting can provide a basis for low-carbon economic dispatch of power systems with a high proportion of renewable energy. Regression networks with many times training based on multi-group multi-configuration still cannot resist the randomness of training processes, resulting in the accuracy of photovoltaic power prediction needs to be improved. This work proposes a weighted fully-connected regression network, including a feature input layer, deep fully-connected layers, particle swarm optimization, and a regression output layer. The proposed model automatically selects two networks from multi-group multi-configuration well-trained regression networks to effectively reduce photovoltaic power prediction errors without additional sensors and data sources. The errors of these two chosen well-trained networks exactly neutralize each other by fixed and simple weights. The results under the one-day-ahead hourly photovoltaic power forecasting of Natal of Brazil show that the proposed method can reduce photovoltaic power prediction errors with at least 75.9954% smaller mean absolute error than the state-of-art methods and 68.2937% than other 18 famous convolutional neural networks methods.