Electricity load forecasting remains a critical function for grid operators seeking to maintain reliable and efficient power system operations. Traditional forecasting methods often struggle to extract meaningful patterns from complex, multi-dimensional data sources such as temperature, humidity, historical demand, and temporal variations. Researchers have now proposed a dual-branch neural network architecture designed to overcome these limitations through advanced machine learning techniques.
The proposed model employs two parallel processing channels: one branch combines an Effective Channel Attention Network with Temporal Convolutional Networks to capture sequential dependencies in load data, while a second branch uses an Informer architecture for alternative pattern recognition. This parallel approach allows the system to evaluate multiple feature extraction strategies simultaneously, improving overall prediction robustness. Additionally, elastic net regression pre-processes high-dimensional meteorological inputs, reducing noise and irrelevant variables before entering the neural networks.
Testing on real-world datasets revealed substantial performance gains. The model achieved a root mean square error of 170.65 kilowatts, mean absolute error of 17.35 percent, and R-squared coefficient of 0.9732—metrics that significantly exceed established benchmarks including LSTM, BiLSTM, Transformer, and Informer standalone models. Compared to conventional LSTM networks, the new approach reduced forecasting errors by 57.9 percent while maintaining stable performance across varying computational batch sizes.
Ablation studies confirmed that both the attention mechanism and dual-branch architecture contributed meaningfully to performance improvements, suggesting neither component was redundant. The model's superior generalization capabilities indicate it could reliably support grid operations under diverse load conditions and system configurations. As renewable energy penetration increases grid variability, such improved forecasting tools may prove invaluable for dispatch optimization, reserve scheduling, and demand-response coordination.
