🤖 machine learning

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

This paper introduces LoadKAN, a novel interpretable hybrid framework that integrates a feature-isolated temporal attention mechanism with Kolmogorov-Arnold Networks to achieve competitive electricity load forecasting accuracy while enabling granular analysis of complex, non-linear relationships between human mobility patterns and energy demand.

Original authors: Jinhao Li, Hao Wang

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jinhao Li, Hao Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Predicting Power with a "Glass Box"

Imagine you are trying to predict how much electricity a city will use tomorrow. You need to be accurate so the power grid doesn't crash, but you also need to understand why the prediction is what it is.

Most modern AI models are like black boxes. You put data in, and a prediction comes out, but no one knows exactly how the machine made that decision. It's like a chef who cooks a delicious meal but refuses to tell you the recipe or what ingredients they used.

This paper introduces a new model called LoadKAN. Think of LoadKAN as a "Glass Box" chef. It not only cooks a delicious meal (makes an accurate prediction) but also lays out every ingredient and shows you exactly how they interact to create the final flavor. It is built on a new type of AI called a Kolmogorov-Arnold Network (KAN), which is designed to be transparent by default.

The Problem: Why Old Models Struggle

The researchers noticed two main issues with existing AI models for electricity:

They are opaque: You can't easily see how factors like weather or human movement affect the power grid.
They get confused by time: Electricity data changes constantly over time (like a heartbeat). Standard AI models often get "jumbled" when trying to track these complex rhythms, especially when you add new data like human movement.

The Solution: LoadKAN's Two-Step Recipe

The authors built LoadKAN with two specific stages, which they describe as a "Feature-Isolated Temporal Attention Mechanism" combined with a KAN module. Here is the simple version:

Step 1: The "Soloist" Stage (Feature-Isolated Attention)

Imagine a choir where every singer is trying to sing at once. Usually, the loud voices drown out the quiet ones, and the song becomes a mess.
LoadKAN does something different. It puts each singer (each data point like "weather," "traffic," or "people at home") in their own soundproof booth.

What it does: It listens to the "Weather" singer alone to understand its rhythm. Then it listens to the "Human Mobility" singer alone.
Why: This prevents the data from getting "entangled." It ensures that the model learns the unique story of each factor before mixing them together.

Step 2: The "Transparent Mixer" Stage (The KAN Module)

Once the soloists have practiced, they come together to sing the final song (the electricity prediction).

The Magic: Instead of a black box mixing the sounds, LoadKAN uses a special "mixer" (the KAN) that draws a visible map of how each voice changes the song.
The Result: You can look at the map and say, "Ah, when people stay home more, the power usage goes down," or "When people go to parks, usage goes up." The model literally draws the curve of these relationships, making it easy to read.

The Secret Ingredient: Human Mobility

The researchers tested this model using data from three major US power markets: New York (NYISO), California (CAISO), and Texas (ERCOT). They added a special ingredient: Human Mobility Data.

They tracked six types of movement:

Retail & Recreation (restaurants, movies)
Grocery & Pharmacy (shopping)
Parks
Transit Stations (subways, buses)
Workplaces
Residential (people at home)

The Findings:

Accuracy: LoadKAN was just as good at predicting electricity as the most advanced "black box" models. In Texas, it was actually the best at predicting.
The "Mobility" Boost: Adding data about where people were moving made every model better, but LoadKAN used this data most effectively.
Market Differences: The model revealed that different cities react differently to movement.
- In New York: When people go to Transit Stations, electricity usage actually drops. The model figured out that if people are on the subway, they aren't at home or in offices using power.
- In Texas: Parks were a huge factor. When people went to parks, power usage changed significantly, likely due to air conditioning in recreational areas or travel.
- In California: Workplaces and Grocery stores had a very complex, shifting relationship with power usage that changed day by day.

Why This Matters

The paper claims that LoadKAN solves a major problem: Trust.

In the past, if you wanted to know why a model predicted a power surge, you had to use a separate, clumsy tool to guess. LoadKAN builds the explanation directly into the math. It allows power grid operators to see the "recipe" of the future. They can see exactly how a rainy Tuesday or a holiday weekend with people at home will impact the grid, without having to guess what the AI is thinking.

In short: LoadKAN is a highly accurate electricity predictor that doesn't hide its work. It separates the data to understand each piece individually, then mixes them together in a way that lets us see exactly how human behavior drives the lights on and off.

Technical Summary: Interpretable Kolmogorov-Arnold Network with Feature-Isolated Temporal Attention for Electricity Load Forecasting

Problem Statement
Accurate electricity load forecasting is fundamental to power system stability, efficiency, and economic viability. While deep learning models (e.g., Transformers, LSTMs) have achieved high predictive accuracy, they typically operate as "black boxes," lacking the interpretability required to understand the specific drivers of forecast outcomes. This opacity hinders the ability of system operators to grasp how complex exogenous factors, particularly human mobility patterns, influence electricity consumption. Furthermore, the direct application of the emerging Kolmogorov-Arnold Network (KAN) to time-series forecasting faces challenges, including training instability and an inability to effectively model complex temporal dependencies without specific architectural adaptations. Existing hybrid approaches that merely replace Multi-Layer Perceptron (MLP) layers with KAN modules often fail to fully leverage KAN's inherent interpretability, as they do not preserve the distinct dynamics of individual input features.

Methodology: The LoadKAN Framework
To address these gaps, the authors propose LoadKAN, a novel hybrid and interpretable framework designed for electricity load forecasting. The architecture consists of two synergistic stages:

Feature-Isolated Temporal Attention Mechanism:
- Objective: To process each input feature sequence (e.g., historical load, weather, and six distinct human mobility metrics) independently before any cross-feature interaction occurs.
- Mechanism: The input batch is transposed and projected into a higher-dimensional embedding space via 1D convolutions. Crucially, the model employs a feature-wise Multi-Head Self-Attention (MHSA) mechanism. Unlike standard architectures that entangle all features into latent embeddings, this stage treats each feature's temporal history as an independent sequence.
- Aggregation: Temporal information for each feature is aggregated by selecting the output of the last time step, followed by a feature-isolated 1D convolution and a hyperbolic tangent ( $\tanh$ ) activation. This produces a distilled, temporally-aware scalar representation for each feature, ensuring that the unique temporal dynamics of variables like "Retail & Recreation" or "Residential" are preserved without signal dominance or noise propagation from other features.
Interpretable KAN Module:
- Objective: To map the processed feature representations to the future load forecast while maintaining structural transparency.
- Mechanism: The distilled features are fed into a KAN layer constructed with learnable B-spline activation functions on the edges. Based on the Kolmogorov-Arnold Representation Theorem, the model approximates the multivariate function as a sum of univariate functions: $\hat{y} = \sum f_m(r_m) + \beta$ .
- Interpretability: Because the input to the KAN layer consists of isolated feature representations, the learned spline functions explicitly model the non-linear relationship between a specific real-world feature (e.g., mobility to parks) and the electricity load. This allows for direct visualization of functional relationships (e.g., saturation points, threshold effects) without relying on post-hoc explanation tools like SHAP or LIME.

Key Contributions

Architectural Innovation: The paper introduces LoadKAN, which uniquely integrates a feature-isolated temporal attention mechanism with a KAN module. This design prevents the entanglement of feature dynamics, a prerequisite for meaningful interpretability in time-series forecasting.
Empirical Validation: The framework was evaluated on datasets from three representative U.S. electricity markets: NYISO, CAISO, and ERCOT. The study demonstrates that LoadKAN achieves competitive performance against extensively tuned state-of-the-art black-box benchmarks (including Informer, TCN, and Transformer), often securing top-tier results (e.g., best performance in ERCOT).
Ablation on Mobility Features: The study quantifies the impact of human mobility data, showing that integrating these features significantly improves forecasting accuracy across all models, with LoadKAN demonstrating a strong aptitude for leveraging this supplementary information.
Granular Interpretability: The paper provides a detailed analysis of the learned non-linear relationships between six mobility patterns and electricity load. It reveals market-specific dependencies, such as the varying influence of transit usage in NYISO versus commercial mobility in CAISO, and the dominant sensitivity of park-related mobility in ERCOT and NYISO.

Results

Performance: LoadKAN consistently performs within the top tier of benchmarks. In the ERCOT market, it achieved the best overall performance (MAPE: 2.91%, $R^2$ : 0.9429), outperforming the best baseline (Transformer). In NYISO and CAISO, it remained highly competitive, with MAPE values within 0.17% and 0.13% of the best-performing models, respectively, while offering superior interpretability.
Mobility Impact: The inclusion of human mobility features resulted in an average reduction of MAPE by approximately 16.28% and RMSE by 16.90% across all models and markets.
Sensitivity Analysis: Quantitative sensitivity analysis of the learned activation functions revealed distinct regional behaviors. For instance, "Parks" mobility exhibited the highest sensitivity in NYISO and ERCOT, while commercial features (Workplaces, Retail) showed higher sensitivity in CAISO. The analysis also uncovered complex, non-linear, and sometimes counter-intuitive relationships, such as increased residential presence correlating with reduced system load in certain markets, likely due to the substitution of energy-intensive commercial activities.

Significance and Claims
The paper claims that LoadKAN successfully bridges the gap between the predictive power of deep learning and the structural interpretability required for informed decision-making in energy systems. By avoiding the "black box" nature of standard neural networks and the limitations of pure KAN structures on time-series data, LoadKAN offers a transparent mechanism to understand how specific drivers, particularly human mobility, influence electricity demand. The authors emphasize that their approach allows stakeholders to visualize and quantify the non-linear dependencies between input features and load forecasts, providing insights that are often obscured by opaque models. This capability is presented as a crucial enabler for understanding the impact of dynamic exogenous features on electricity consumption, supporting more informed system operations in an evolving energy landscape.