Explainable AI based System for Supply Air Temperature Forecast

TL;DR

The paper tackles explainability in ASAT forecasting for HVAC control by applying Shapley-value explanations to a regression-based forecast, with a focus on transparency through local, contrastive slices. It models ASAT using a sliding-window time-series approach that incorporates ambient temperature, room-temperature average, and ASAT history encoded in a 2D dynamic array with a lag of 37, aligning with Iterative Learning Control concepts, and trains a linear model with $Huber$ loss. The key contributions include demonstrating SHAP-based local explanations for each forecast point, identifying the most influential features (notably end-of-day ASAT history and RT-avg) while noting lower influence from ambient temperature, and providing a practical mechanism to justify curve changes to customers. The findings enhance transparency and interpretability in HVAC control, enabling safer and more justifiable decision-making in high-stakes environments; the method is grounded in the SHAP framework, with the fundamental property $\sum_i \phi_i = \hat{f}(x) - E[\hat{f}(X)]$ ensuring fair attribution of prediction deviations across features.

Abstract

This paper explores the application of Explainable AI (XAI) techniques to improve the transparency and understanding of predictive models in control of automated supply air temperature (ASAT) of Air Handling Unit (AHU). The study focuses on forecasting of ASAT using a linear regression with Huber loss. However, having only a control curve without semantic and/or physical explanation is often not enough. The present study employs one of the XAI methods: Shapley values, which allows to reveal the reasoning and highlight the contribution of each feature to the final ASAT forecast. In comparison to other XAI methods, Shapley values have solid mathematical background, resulting in interpretation transparency. The study demonstrates the contrastive explanations--slices, for each control value of ASAT, which makes it possible to give the client objective justifications for curve changes.

Figures (7)

• Figure 1: Dropping of the time step column and shifting of the measures column.
• Figure 2: Configuration and encoding of the feature vector used in modelling of ASAT.
• Figure 3: Example of a building a 2D Dynamic array of ASAT history based on two previous days. 2D array is given in time-step representation.
• Figure 4: The forecasted and true ASAT control values, as well as the difference between the control curves.
• Figure 5: Distribution of model parameters--coefficients, of models trained to forecast ASAT control values. ASAT history data in time-step representation.