TimeGPT vs. LSTM-CNN for Photovoltaic Energy Forecasting
Year: 2026
Type: conference-paper
Abstract: The large-scale deployment of renewable energy sources at the grid edge has significantly increased the complexity of energy management. In this context, accurate solar energy forecasting is essential, yet challenging due to the intermittent and stochastic nature of photovoltaic generation. Recently, foundation time-series models such as TimeGPT have emerged as a promising alternative paradigm to deep learning approaches. In this paper, we compare the PV energy production forecasting performance of a hybrid deep learning model combining Long Short-Term Memory and Convolutional Neural Network (LSTM-CNN) and TimeGPT. We define an evaluation methodology based on two different input history lengths: a long-term history of one year and a short-term history of one month. For each model, forecasts are generated for three prediction horizons: 24, 72 and 120 hours ahead. This allows for analyzing the models performance under varying temporal contexts, allowing to determine both short and long terms dependencies. Results show that even with extended historical context, TimeGPT performs better in short-term forecasting (24 hours), while LSTM-CNN achieves higher accuracy for medium horizons of (72 / 120 hours) due to explicit sequential modeling.
Institution: Technical University of Cluj-Napoca
Language: English
Field: Energy
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7214627353