From Grid History to Future Clouds : Temporal convolution Neural Network Driven Horizon-Aware Solar Power Forecasting with Fingrid and FMI Data

Abstract

This thesis investigated fast and reliable short-term forecasting of solar power to support rising demand and decarbonization targets. Using data from the Finnish national electricity transmission grid operator (Fingrid) and the Finnish Meteorological Institute (FMI), the study described the relationship between weather and generation and proposed a compact temporal convolutional network that is practical for deployment. The inputs were combined from standard meteorological variables, cyclic calendar encodings, and exogenous power system series such as wind and hydro production to provide physically meaningful signals. Evaluation was conducted with chronological splits across seasons and time of day groups and compared different horizons against strong baselines. Feature importance was quantified through permutation tests to support interpretability and operational trust.

The results showed that the temporal convolutional network captured diurnal structure and major ramps while remaining competitive with short-horizon baselines. Hour-of-day encodings and related diurnal terms were identified as the most influential drivers, followed by UV index and other irradiance proxies. Their contribution typically increased at higher horizons, reflecting the growing value of near future cloud-cues beyond the instantaneous state. Wind and hydro production series provided complementary information, especially at night and during transition periods, by signalling weather regimes that also affect solar conditions. The study highlighted practical opportunities for improvement, including the use of forecast-aligned meteorology for multi-hour horizons, the incorporation of direct cloud nowcasts, and the weighting of ramp errors in the loss function to emphasize operationally critical moments. While the analysis was limited by national-level aggregation which will require further fine-grained adaptation for site-level deployment, the findings indicated that a compact and interpretable temporal convolutional network can support virtual power plant optimization and real-time decision making, with clear routes for enhancement through richer sensing and improved objective design.