New model for day-ahead solar forecasting in areas with limited data

Scientists from South Korea’s Inha University have developed a novel solar irradiance model that can purportedly predict data for a whole year with only two weeks of learning period.

The model is based on a reinforcement learning algorithm and is claimed to be particularly suitable with limited accumulated data.

“Reinforcement learning is an algorithm that independently maximizes rewards through trial-and-error within a given environment,” the academics explained. “This is the first attempt to use a reinforcement learning for solar irradiance prediction model.”

In the study “Solar irradiance prediction using reinforcement learning pre-trained with limited historical data,” published in Energy Reports, the researchers explained that the learning data was set for the first two weeks of the year, while the prediction period was set for the remaining part of the year.

As inputs for those two weeks, the scientists used real data from Cape Town, South Africa, based on four metrics – sky cover, temperature, humidity, and out-of-atmosphere solar irradiance. They then compared the analysis made by the model with that of two reference models based on long short-term memory (LSTM), which is a kind of recurrent neural network capable of learning order dependence in sequence prediction problems.

“The LSTM model is a type of recurrent neural network (RNN) deep learning,” the research group added. “It introduces a memory cell state to the RNN. The memory cell state demonstrates superior performance in managing long sequence inputs.”

On a monthly average, the researchers found that the proposed model had a prediction error 0f 31.5 W/m2, which represented a 7% variation from the actual data. In comparison, the two LSTM models used in the experiment showed an average error of 57.4 W/m2 or 12.8% and of 40.5 W/m2 or 9.2%, respectively.

Furthermore, the researchers ascertained that, with the proposed model, approximately 82.8% of the points were distributed within 10% of the upper and lower errors of the actual results. With the two LSTM models, these values decreased to 71.9% and 78.8%, respectively.

“The proposed model displayed more optimized performance with only two weeks of learning data when compared to previous studies which typically required several years’ worth of weather data for learning,” the researchers concluded. “In particular, the long-term performance of the model was improved by learning the out-of-atmosphere solar irradiance as an input.”