Transfer learning strategies for solar power forecasting under data scarcity

Abstract

Accurately forecasting solar plants production is critical for balancing supply and demand and for scheduling distribution networks operation in the context of inclusive smart cities and energy communities. However, the problem becomes more demanding, when there is insufficient amount of data to adequately train forecasting models, due to plants being recently installed or because of lack of smart-meters. Transfer learning (TL) offers the capability of transferring knowledge from the source domain to different target domains to resolve related problems. This study uses the stacked Long Short-Term Memory (LSTM) model with three TL strategies to provide accurate solar plant production forecasts. TL is exploited both for weight initialization of the LSTM model and for feature extraction, using different freezing approaches. The presented TL strategies are compared to the conventional non-TL model, as well as to the smart persistence model, at forecasting the hourly production of 6 solar plants. Results indicate that TL models significantly outperform the conventional one, achieving 12.6% accuracy improvement in terms of RMSE and 16.3% in terms of forecast skill index with 1 year of training data. The gap between the two approaches becomes even bigger when fewer training data are available (especially in the case of a 3-month training set), breaking new ground in power production forecasting of newly installed solar plants and rendering TL a reliable tool in the hands of self-producers towards the ultimate goal of energy balancing and demand response management from an early stage.

Figure 1

A map depicting the locations of the inspected PV plants. The 7 PV plants are located in four different Portuguese cities (Lisbon, Setubal, Faro, Braga) allowing for assessment of model performance regardless of the city. The image on the right offers a focus on the city of Lisbon where 4 PV plants are located (map created with https://www.mapcustomizer.com/, OpenStreetMap contributors).

Figure 2

The transfer learning process.

Figure 3

The long short-term memory (LSTM) model architecture.

Figure 4

The power output expected under clear sky conditions for the base PV as a function of the hour of the day (ranging between 0 and 23) and the day of the year (ranging between 0 and 255).

Figure 5

Example illustrating how the solar power forecasting model performs in comparison with the smart persistence model. The horizontal axis indicates the hourly time-step of the evaluation period, while the vertical axis shows the solar power production. The example refers to a randomly selected summer week (between hours 54 and 222 of the validation set, corresponding to 04-08-2020 and 11-08-2020) and a randomly selected winter week (between hours 4060 and 4228 of the validation set, corresponding to 18-01-2021 and 25-01-2021) of the evaluation period.

Figure 6

Boxplot that summarizes the performance of the four stacked LSTM models for the six target PV plants based on the RMSE. Base stands for the model that no TL has been applied, while TL1, TL2, TL3 stand for the three TL strategies, respectively.

Figure 7

Comparative performance of the four models based on RMSE index for 3 month, 6 month, 9 month and 12 month training period for the six target domain PVs.