Short-Term Photovoltaic Power Forecasting Using an LSTM Neural Network and Synthetic Weather Forecast

In this paper, a forecasting algorithm is proposed to predict photovoltaic (PV) power generation using a long short term memory (LSTM) neural network (NN). A synthetic weather forecast is created for the targeted PV plant location by integrating the statistical knowledge of historical solar irradiance data with the publicly available type of sky forecast of the host city. To achieve this, a K-means algorithm is used to classify the historical irradiance data into dynamic type of sky groups that vary from hour to hour in the same season. In other words, the types of sky are defined for each hour uniquely using different levels of irradiance based on the hour of the day and the season. This can mitigate the performance limitations of using fixed type of sky categories by translating them into dynamic and numerical irradiance forecast using historical irradiance data. The proposed synthetic weather forecast is proved to embed the statistical features of the historical weather data, which results in a significant improvement in the forecasting accuracy. The performance of the proposed model is investigated using different intraday horizon lengths in different seasons. It is shown that using the synthetic irradiance forecast can achieve up to 33% improvement in accuracy in comparison to that when an hourly categorical type of sky forecast is used, and up to 44.6% in comparison to that when a daily type of sky forecast is used. This highlights the significance of utilizing the proposed synthetic forecast, and promote a more efficient utilization of the publicly available type of sky forecast to achieve a more reliable PV generation prediction. Moreover, the superiority of the LSTM NN with the proposed features is verified by investigating other machine learning engines, namely the recurrent neural network (RNN), the generalized regression neural network (GRNN) and the extreme learning machine (ELM).