Convolution-Based Sequence to Sequence Model for the Next-Day Photovoltaic Power Forecasting

In photovoltaic power forecasting, problems such as seasonal characteristics change and insufficient historical data are often faced. And because of the randomness of actual weather changes, it is very difficult to forecast the power curve of the next day. In this paper, we propose a series of solutions to these difficulties. By the time series decomposition, we solve the problem of data distribution fluctuation caused by seasonal feature changes. By using the convolution-based sequence to sequence model, we effectively combined historical observed data with future forecasted weather data, and accurately learn the time correlation. We adopt the multi-kernel gated convolutional neural networks to achieve learning of accurate local cross features. By analyzing the correlation between the features and the actual power at different times of one day, we propose a time-specific forecasting layer. Experiments show that our method has significant performance improvement compared with baseline models.