Very short-term production prediction for photovoltaic plants using Temporal Convolutional Networks

Abstract

Very short-term forecasting of solar photovoltaic energy production at national scale is challenging due to the high variability and spatial aggregation of generation across large territories. This paper evaluates Temporal Convolutional Networks (TCN) — a deep learning architecture based on causal and dilated convolutions — for nowcasting national-level solar production at one-hour and fifteen-minute horizons, using data from Spain sourced from the European Network of Transmission System Operators for Electricity. Multivariate models augmented with past weather observations (solar irradiance and sun height) are compared against linear regression baselines. Results demonstrate that the multivariate TCN substantially outperforms linear regression at the one-hour horizon, and achieves consistent improvement at the fifteen-minute horizon. The relative contribution of architecture and weather features is resolution-dependent: at hourly granularity, the TCN architecture itself provides the dominant gain, while at the fifteen-minute scale the inclusion of weather covariates becomes the primary driver of accuracy, reflecting the greater atmospheric variability at finer temporal scales. A key finding is that past-only weather inputs are sufficient for accurate nowcasting, eliminating the need for future meteorological forecasts as model inputs. The results support the practical applicability of TCN-based models for national-level solar energy integration and provide a data-driven feature-selection criterion for similar renewable energy forecasting tasks.