Utility-Scale Solar Power Forecasting: Single-Step vs. Multi-Step Feed Forward Neural Networks

Photovoltaic (PV) power forecasting serves as an important tool to manage the intermittent nature of PV-system power supply. In recent years, machine learning based forecast models have proven to be an effective solution to reduce the uncertainties associated with PV power. Typically, from a time-series forecasting perspective, practitioners either deploy a single-step or multi-step forecasting methodology. Single-step forecast models predict a value for a single point in time, whereas multi-step models predict multiple future values at once for a range of time. From literature, it is not evident which of these two approaches are more advantageous for an intra-day (1–6 h) PV power forecasting scenario. Given the high computational cost of single-step forecasting solutions, compared to multi-step solutions, it is important to understand the difference in performance of these two approaches. Results are generated for a real-world forecasting scenario, where the well-established Feedforward Neural Network (FFNN) model is used to deliver forecasts for a 75 MW rated PV system. From the overall results, both single-step and multi-step FFNNs delivered forecasting errors within 2 % of one another, with single-step models proven to have an improved ability to capture the complex PV system dynamics. However, from an engineering perspective, it is concluded that multi-step forecasting is a more effective forecasting solution when the available processing power and time is limited. The findings of this study have implications for energy producers and practitioners in the field of solar PV forecasting.