Interpretable Kolmogorov-Arnold Network with Feature-Isolated Temporal Attention Mechanism for Electricity Load Forecasting

Electricity load forecasting is essential for power system stability, but modern deep learning models often function as opaque black boxes. While Kolmogorov-Arnold Networks (KAN) offer a promising alternative due to their learnable spline-based activation functions, they struggle with complex temporal patterns when applied directly to time-series data. This paper introduces LoadKAN, a hybrid framework designed to bridge this gap.

LoadKAN employs a two-stage architecture. First, it uses a feature-isolated temporal attention mechanism to extract dynamics from each input feature (e.g., historical load, weather, and six categories of human mobility) independently. This prevents feature entanglement, ensuring that the temporal patterns of each variable are distilled before being passed to the second stage. The second stage utilizes a KAN module, which replaces traditional fixed-activation neurons with learnable splines on the network edges, allowing for explicit modeling of non-linear relationships between inputs and the final load forecast.

LoadKAN was evaluated using datasets from three major U.S. electricity markets: NYISO, CAISO, and ERCOT. The results demonstrate that LoadKAN maintains highly competitive predictive accuracy compared to extensively-tuned, state-of-the-art black-box deep learning benchmarks.

Beyond accuracy, the framework provides significant interpretability advantages. By analyzing the learned spline functions, the authors performed a granular sensitivity analysis of how different human mobility patterns—such as workplace presence versus retail activity—influence electricity consumption. The model successfully identified market-specific dependencies, revealing how the impact of human behavior on energy demand varies across different geographical regions. This transparency allows system operators to move beyond simple point forecasts and gain actionable insights into the underlying drivers of electricity consumption.

As power grids integrate more renewable energy and face changing demand patterns from electric vehicles and shifting human activities, understanding the 'why' behind a forecast is as important as the 'what.' LoadKAN demonstrates that it is possible to achieve high-performance forecasting without sacrificing transparency. By embedding interpretability directly into the neural architecture, this approach reduces reliance on potentially unstable post-hoc explanation methods, providing a more reliable tool for energy management and grid planning.