Attention-based deep networks are increasingly proposed for day-ahead load forecasting in smart grids. This study develops a CNN-LSTM forecaster whose temporal attention query is conditioned on the calendar type and hourly temperature of the target day, and it evaluates the model under a protocol designed to avoid three common biases: target leakage across data splits, unequal inputs for benchmarks, and comparison of seed ensembles with single runs. Hourly PJM East load and temperature data for 2013–2018 were split chronologically with purging, and the model was compared with persistence, ridge regression, a multilayer perceptron (MLP), an LSTM and a CNN-LSTM, each stochastic model being trained with three random seeds. With temperature inputs, the proposed model reached a test MAPE of 2.21%, was the most accurate recurrent model, and its attention layer reduced the error of the CNN-LSTM by 20.8% (p < 0.001). However, an MLP given the same information was equally accurate in single runs, significantly more accurate as a seed ensemble (2.13%, p = 0.018) and more robust to temperature-forecast errors, and an exploratory combination of both models reached 2.02%. Deep hybrid forecasters should therefore be judged against well-configured feed-forward benchmarks under symmetric, leakage-free protocols.
Copyrights © 2026