Building LSTM-Based Model for Solar Energy Forecasting

Forecasting of solar energy generation is often based on estimating the radiation received from the sun, known as Global Horizontal Irradiance (GHI), which includes both direct and diffused radiation. Various models are used for forecasting, including physical models, statistical models, machine learning models, and deep learning-based models. The deep learning-based model, Long Short-Term Memory (LSTM), is particularly suited for understanding complex and non-linear patterns in time series data. LSTM models have a three-dimensional shape, with the number of input features, number of timesteps, and batch size being key parameters. The number of input features determines the variables used for prediction, while the number of timesteps, or input window size, determines the number of past values used for prediction. The batch size parameter controls the size of the gradient descent used for training the model. One of the benefits of LSTM is its ability to handle time series data without the need for pre-processing, such as removing trends and seasonality. However, research has shown that pre-processing can improve the performance of LSTM models. In one experiment, the performance of LSTM trained on raw time-series data was compared to that trained on pre-processed data. The results showed that LSTM trained on raw time-series data gave better results, as measured by normalized RMSE and Explained Variance Score. Another design question addressed in the research was whether treating time steps as separate features or as part of a time series setup is more effective. The results showed that treating time steps as part of a time series setup gave better results. This is because LSTM models are designed to capture temporal relationships in data, and treating time steps as separate features can lead to a loss of this information. The research also investigated the relationship between the number of nodes in an LSTM network and its performance. The intuition is that increasing the number of nodes will always improve performance, but this is not necessarily the case. The results showed that for simple scenarios, a small number of nodes is sufficient, while for more complex scenarios, a larger number of nodes is required. However, beyond a certain point, adding more nodes does not necessarily improve performance, and may even lead to overfitting. In conclusion, LSTM can be a good model for solar forecasting, but it requires careful consideration of the input data and the structure of the network. Raw time series data should be used, and time steps should be treated as part of a time series setup. The number of nodes in the network should be chosen based on the complexity of the scenario, and more complex scenarios will require more complex structures.