Abstract
<jats:p>This paper investigates the application of the Temporal Fusion Transformer (TFT) model for multi-horizon forecasting of electricity generation from photovoltaic power plants (PV systems). Accurate prediction of solar generation is essential for reliable operation of modern power systems due to the variability and uncertainty of renewable energy sources. TFT is a deep learning architecture designed for time series analysis, combining variable selection mechanisms and attention layers to capture both short-term and long-term dependencies. The model performs forecasting for multiple horizons (1, 6, and 24 hours ahead) and provides interpretable results by estimating the contribution of input features. The model uses three types of input data: historical observations (power generation and weather parameters such as solar irradiance, temperature, and wind speed), known future inputs (time and solar position), and static characteristics of the PV system. Training and evaluation are conducted using publicly available datasets from NREL (PVDAQ and NSRDB), which include real-world measurements. The proposed approach is compared with baseline models such as LSTM, GRU, and ARIMA using standard metrics (MAE, RMSE, MAPE). The results demonstrate that TFT provides higher accuracy across all forecasting horizons, especially for longer-term predictions. Additionally, the model generates prediction intervals, allowing uncertainty estimation, which is important for energy system planning. The study focuses on integrating forecasting models into intelligent monitoring and control systems for PV plants based on multi-agent architectures. The proposed approach enables automated data processing and supports adaptive decision-making.</jats:p>