Towards Energy Efficiency: Forecasting Indoor Temperature via Multivariate Analysis

Technical Summary: Towards Energy Efficiency: Forecasting Indoor Temperature via Multivariate Analysis

Problem Statement
The paper addresses the challenge of improving energy efficiency in residential buildings, specifically targeting Heating, Ventilation, and Air Conditioning (HVAC) systems, which constitute a significant portion of residential energy consumption (53.89% in the studied SMLsystem house). The core problem is the need to predict indoor temperature accurately to enable proactive HVAC control. The authors note that maintaining a temperature is more energy-efficient than cooling or heating it from a different state due to thermal inertia. Therefore, forecasting future indoor temperatures allows the building automation system to anticipate thermal conditions and adjust HVAC operations (e.g., activating a heat pump) to maintain comfort while minimizing energy usage. The study focuses on the "Small Medium Large System" (SMLsystem), a solar-powered prototype house built for the Solar Decathlon Europe 2012 competition.

Methodology
The research employs a multivariate time series forecasting approach using Artificial Neural Networks (ANNs) and compares them against standard statistical methods.

• Data Acquisition and Preprocessing: Data was collected from the SMLsystem house using a KNX-based monitoring framework with 88 sensors and 49 actuators. The raw data, sampled at 1-minute intervals, was sub-sampled to 15-minute intervals by computing means. The dataset was partitioned into training (21 days, March 2011), validation (7 days, March 2011), and test (7 days, June 2011) sets.
• Input Variables (Covariates): The models utilized the following input features:
  • Indoor temperature ($d$): The target variable, differentiated to improve generalization.
  • Hour of the day ($h$): Encoded as a locally-encoded category (24 components) to handle the cyclical nature of time.
  • Sun irradiance ($W$), Indoor relative humidity ($H$), CO2 levels ($Q$), and Rain status ($R$): Normalized using mean and standard deviation.
• Forecasting Models:
  • Statistical Baselines: Exponential Smoothing (ETS) and Auto-Regressive Integrated Moving Average (ARIMA) models were used as benchmarks.
  • Artificial Neural Networks (ANNs): Multi-layer perceptrons were trained using the backpropagation algorithm with momentum and L2 regularization (weight decay). The study utilized a Multiple Input Multiple Output (MIMO) approach to predict a future window of size $Z=12$ (3 hours) in a single step.
  • Ensemble Strategies: To improve robustness, the authors combined multiple ANN models trained with different input window sizes ($I(x)$). Two combination schemes were tested:
    • COMB-EQ: Uniform linear combination (equal weights).
    • COMB-EXP: Exponential linear combination (softmax weights) based on the inverse of the Mean Absolute Error (MAE) on the validation set.
• Evaluation: Performance was measured using Symmetric Mean Absolute Percentage Error (SMAPE), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) on the validation and test partitions.

Key Results

• ANN Superiority: ANNs significantly outperformed standard statistical methods (ETS and ARIMA) with a confidence level greater than 99%.
• Optimal Covariates: The most effective model configuration combined indoor temperature with the hour of the day and sun irradiance ($d + h + W$). Adding further covariates (humidity, CO2, rain) yielded only marginal improvements.
• Ensemble Performance: The COMB-EXP (exponential combination) strategy generally produced slightly better results than the uniform (COMB-EQ) or single-best (BEST) models, though the differences were often not statistically significant.
• Forecast Accuracy: The best-performing model (COMB-EXP with $d + h + W$) achieved a SMAPE of approximately 0.45% and an MAE of roughly 0.11°C on the test partition.
• Horizon Sensitivity: Error rates increased as the forecast horizon extended, and confidence intervals widened. However, the inclusion of covariates like sun irradiance became more critical for longer horizons (greater than 90 minutes).
• Mutual Information (MI): An MI analysis confirmed that sun irradiance has a strong relationship with indoor temperature, particularly during the day. Interestingly, humidity and air quality showed high MI values, but their inclusion in the model did not significantly enhance performance beyond what was provided by time and irradiance data.

Significance and Claims
The paper claims that accurate indoor temperature forecasting is a viable strategy for energy-efficient HVAC control. By predicting future thermal states, the SMLsystem can adapt its operations to maintain comfort with reduced energy consumption. The study demonstrates that multivariate ANNs, particularly when combined with specific environmental covariates (hour and irradiance), offer a robust solution for this task, outperforming traditional statistical time series methods.

The authors emphasize that their work provides a practical framework for integrating forecasting into home automation systems (Ambient Intelligence) to balance comfort and energy savings. They modestly note that the study is based on data from a single house in Southern Europe over a limited timeframe, suggesting that future work should validate these findings across different seasons, locations, and building types. The ultimate goal is to utilize these predictive capabilities to design predictive control systems that optimize energy subsystems beyond just temperature regulation.