Embedded Deep Learning for Short-Term PV Forecasting Under Export Constraints

The increasing penetration of photovoltaic (PV) systems requires accurate and stable short-term forecasting to ensure reliable grid operation under operational constraints. This paper investigates short-horizon multi-step PV power forecasting using one full year of high-resolution (5 min) real-world data from a 111-kW grid-connected rooftop installation. The forecasting problem is formulated as a direct multi-output supervised learning task with a 30 min prediction horizon. A comprehensive comparative evaluation is conducted across baseline (persistence), tree-based (XGBoost), and deep learning architectures (LSTM, GRU, and Temporal Convolutional Networks—TCN). Results show that deep learning models significantly outperform conventional baselines, with LSTM achieving the lowest normalized RMSE (≈10.3%), while TCN provides a competitive trade-off between predictive accuracy, temporal stability, and computational efficiency. The direct multi-step formulation was adopted to reduce potential error propagation effects commonly observed in recursive forecasting approaches. Beyond forecasting accuracy, the study evaluates computational complexity and inference latency to assess practical deployability in resource-constrained environments. The proposed framework demonstrates that high-resolution real-world PV forecasting can achieve both strong predictive performance and operational feasibility. These findings contribute to the development of robust short-term forecasting strategies for distributed renewable energy systems operating under regulatory export constraints.