Hybrid nowcasting for solar power plants using satellite-data and Numerical Weather Predictions for (Deep) Machine Learning methods

With the expanding penetration of renewable energy in the energy sector, we observe an ever-increasing need for more accurate weather and production forecasts. They are needed by several energy players: plant owners, system operators, service providers (balancing service providers, energy traders). For the energy market needs, in different countries we can already find almost real-time trading markets; in a likely future scenario, the day-ahead market will disappear in favour of 5-15 minutes ahead market. This trend luckily matches the system operators need of predicting in the very short term the energy fed into the grid, to effectively cope with voltage and congestions problems and manage the ancillary services. Overall, the scenario indicates a compelling need for advanced forecasting techniques.

This article discusses a hybrid solar nowcasting system, predicting energy production from +15 minutes to 3 hours ahead, with a time granularity of 15 minutes. The system combines observed data (especially from satellite) and Numerical Weather Predictions to nowcast data in two steps: the first step is the nowcast of global horizontal irradiance and direct normal irradiance; they are then fed into the following system to predict the energy production. Thus, we disentangle the problem, and we can improve in parallel the two subsystems.

The weather nowcast model core is a Deep Learning method especially suited for time series problems (Long Short Term Memory Network - LSTM). It has been tested over different sites corresponding to different satellite spatial resolution, weather conditions and climate regions. The results are compared with different benchmarks such as the persistence model, smart persistence model and ground truth (where available), obtaining typical annual MAE results over the 15->3 hours between 10 and 80 W/m2. Other metrics (MBE, RMSE, and the forecast score) are calculated to get a deeper view of the results meaning. We also compared results without the availability of NWP (computationally expensive) or ground sensors (not always available in real-time) to understand the benefits of processing those data.

The power production system (fed with the output of the previous model) is a combination of different techniques: Decision trees, KNN, and NN. The performance is typical of 3-6% annual NMAE, depending on the site. We compare the results with the persistence benchmark and we calculate other metrics such as MBE, NRMSE and to get a deeper understanding of the results.

The two-steps model is finally compared with a one-step model only, where just satellite data are fed into a model predicting the power, to compare pros, cons and performance.