Abstract
<title>Abstract</title> <p>Short-term electric load forecasting is essential for the reliable and efficient operation of modern power systems. Although Temporal Convolutional Networks have demonstrated strong capabilities for modeling sequential patterns, their forecasting performance can vary considerably with different hyperparameter settings. This study presents a systematic grid-search-based sensitivity analysis of Temporal Convolutional Network hyperparameters for one-day-ahead electric load forecasting using half-hourly demand data from Queensland, Australia. A reproducible forecasting framework implemented using the NeuralForecast library with a PyTorch backend evaluates 243 configurations generated from three candidate values for each of five hyperparameters: input size, learning rate, encoder hidden size, decoder hidden size, and batch size. All configurations are evaluated under the same chronological data partition and 48-step forecasting horizon using Mean Absolute Percentage Error as the selection criterion. The evaluated configurations produce MAPE values ranging from 1.841% to 5.455%, with an average of 3.499%. The best configuration, consisting of an input size of 336, a learning rate of 0.0005, an encoder hidden size of 64, a decoder hidden size of 128, and a batch size of 16, achieves a MAPE of 1.841%. This result represents a 55.05% reduction in error compared with the reference configuration, which yields a MAPE of 4.096%. The learning rate shows the clearest marginal effect, whereas increasing the number of hidden dimensions does not consistently improve forecasting accuracy. The results demonstrate that effective TCN performance depends on coordinated hyperparameter settings rather than simply increasing model capacity, providing practical guidance for configuring TCN-based short-term load forecasting models.</p>