An Experimental Comparison of TCN, LSTM, and GRU in Short-Term Solar Power Forecasting

Accurate solar power forecasting is a critical factor for stable and efficient power grid operation. Deep learning models, particularly Recurrent Neural Network (RNN) architectures such as Long Short Term Memory (LSTM) and Gate Recurrent unit (GRU), have been widely applied to this problem; however, they still face limitations regarding computational efficiency. This study presents a comprehensive comparative analysis between the Temporal Convolutional Network (TCN) architecture and standard RNN models (LSTM, GRU) for solar power generation forecasting. The models were trained and evaluated using the same real-world dataset. Experimental results show that TCN not only excels in accuracy, achieving the lowest Root Mean Squared Error (RMSE) of 2257.85 kW, but is also significantly more computationally efficient, with a training time nearly four times faster than LSTM. This research validates the TCN as a robust, efficient, and valuable alternative to traditional RNN models for energy forecasting applications.