Short-term solar power forecasting: Investigating the ability of deep learning models to capture low-level utility-scale Photovoltaic system behaviour

Photovoltaic (PV) system power supply is characteristically intermittent. Therefore, PV forecasting is crucial for decision makers responsible for electrical grid stability. With forecast models traditionally trained as macro-level solutions, where a single model emulates the entire PV system, there is uncertainty regarding the ability of these macro-level models to capture the low-level power output dynamics of large multi-megawatt PV systems. Instead, an aggregated inverter-level forecasting methodology is proposed to obtain an enhanced forecasting accuracy. These macro-level and inverter-level forecasting methodologies are implemented with state-of-the-art deep learning based Feedforward neural network, Long Short-Term Memory and Gated Recurrent Unit recurrent neural network models. Results are generated for a real-world scenario, with multi-step forecasts delivered 1-6 h ahead for a 75 MW rated PV system. To ensure the scalability of the proposed methodology, a unique inverter-clustering technique is presented, which reduces the effort of optimising multiple low-level forecast models. A heuristic process of systematic hyperparameter optimisation is also proposed, which serves to guide future forecasting practitioners towards unbiased model development. From the deterministic and probabilistic confidence interval evaluations, overall results demonstrate a marginal increase in forecasting accuracy from the proposed aggregated inverter-level forecasts. The best performing macro-level model obtained Mean Absolute Percentage Error (MAPE) values ranging between 1.42%-8.13% for all weather types and forecast horisons. In comparison, the equivalent inverter-level forecasts delivered MAPE values ranging from 1.27%-8.29%. Finally, it is concluded that deep learning based macro-level forecast models have a sufficient ability to capture low-level PV system behaviour.