Technical Summary: Electricity Consumption Forecasting Using Cooperative Ensemble Learning with SHAP

Problem Statement

Managing electricity consumption in educational institutions presents significant challenges due to the resource's susceptibility to diverse influencing factors, including institutional expansion, fluctuating student numbers, and climatic conditions. In universities, demand is rapidly growing, necessitating efficient energy management for sustainability and operational competitiveness. However, accurate forecasting is complicated by the vulnerability of consumption patterns to equipment malfunctions, grid instabilities, and unique institutional behaviors such as fixed vacation periods, alternating schedules, and sporadic events like conferences.

Existing literature often lacks studies addressing electricity forecasting in educational institutions using cooperative ensemble learning models. Furthermore, traditional time-series models and standard machine learning approaches often struggle with the non-stationary, irregular, and highly variable consumption patterns observed in these settings, particularly when forecasting over extended horizons (e.g., 12 months).

Methodology

The study proposes a novel framework for forecasting electricity consumption 12 months ahead, utilizing data from two campuses of the Federal Institute of Paraná (IFPR): Palmas (Dataset 1) and Coronel Vivida (Dataset 2). The methodology consists of three primary stages:

1. Data Preparation and Feature Selection

• Dataset: Historical consumption data (kWh) from September 2017 to March 2024 (79 months) were analyzed.
• Exogenous Variables: The dataset was augmented with climatic variables (temperature, rain, humidity, wind, solar radiation), a pandemic period identifier (COVID-19), and temporal dummy variables (Month, Year).
• Feature Selection: To reduce complexity and improve efficiency, Shapley Additive Explanations (SHAP) were employed. Based on the importance values derived from XGBoost and Random Forest (RF) models, a sequential feature elimination process was conducted. This reduced the initial set of 47 predictors to 11, retaining the most significant features such as consumption lags, specific temperature metrics, and temporal identifiers.

2. Model Training and Hyperparameter Optimization

Four machine learning models were selected for their proven efficacy: Long Short-Term Memory (LSTM), Random Forest (RF), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGBoost).

• Optimization: Hyperparameters for each model were fine-tuned using two metaheuristic algorithms: Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). The objective was to minimize the Mean Absolute Error (MAE).
• Selection: GA consistently outperformed PSO in finding optimal configurations, particularly for RF and XGBoost, where PSO struggled with convergence. The optimized models were designated as GA-LSTM, GA-RF, GA-SVR, and GA-XGBoost.
• Validation: A hybrid data-splitting approach was used, reserving the final 12 months for testing and employing 5-fold cross-validation for training.

3. The Weaker Separator Booster (WSB) Ensemble

A novel cooperative ensemble learning approach, named Weaker Separator Booster (WSB), was proposed to enhance the performance of the best individual predictor (GA-LSTM).

• Concept: The WSB leverages the "no-free-lunch" theorem by averaging the predictions of the "weaker" models (GA-RF, GA-SVR, GA-XGBoost), which tended to produce linear, stable predictions closer to a horizontal line.
• Mechanism: This average serves as a stable reference point representing "inability to learn" complex patterns. An adaptive boosting mechanism adjusts the GA-LSTM prediction (the strongest learner) by incorporating this reference. The adjustment is governed by a weight parameter that dynamically increases and decreases across the forecast horizon to mitigate error accumulation in recursive forecasting.

Key Results

The proposed approach was evaluated using Mean Absolute Error (MAE) and Symmetric Mean Absolute Percentage Error (sSMAPE).

• Dataset 1 (Palmas Campus):

  • The WSB model achieved the best performance with an sSMAPE of 13.90% and an MAE of 1990.87 kWh.
  • The standalone GA-LSTM model followed with an sSMAPE of 15.35% and MAE of 2264.50 kWh.
  • Both significantly outperformed traditional time-series models (e.g., ARIMA, Holt-Winters) and other neural networks (CNN, ESN).

• Dataset 2 (Coronel Vivida Campus):

  • The WSB model again demonstrated superior performance with an sSMAPE of 18.72% and an MAE of 464.93 kWh.
  • The GA-LSTM model achieved an sSMAPE of 18.80% and MAE of 466.92 kWh.
  • Traditional models and other neural networks showed higher error rates, with sSMAPE values ranging up to 41.59%.

• SHAP Analysis:

  • Feature importance analysis revealed that lagged time-series values (historical consumption) were the dominant predictors for both campuses.
  • Climatic variables had a minimal influence compared to temporal and historical factors.
  • Distinct patterns emerged: In Palmas, lagged values positively influenced forecasts, whereas in Coronel Vivida, they exhibited a negative correlation, suggesting inverse relationships between past and present consumption in that specific location.

Significance and Claims

The paper claims that the proposed Weaker Separator Booster (WSB) approach effectively addresses the complexity of electricity consumption forecasting in educational institutions. Its significance lies in three main contributions:

1. Integration of External Variables: The study systematically incorporates external variables (pandemic, climate, calendar events) and uses SHAP for rigorous feature selection, ensuring models are trained on the most relevant inputs.
2. Comparative Optimization: It provides a comprehensive comparison of diverse AI models (neural networks, ensemble methods, SVMs) optimized via GA and PSO, offering insights into the most suitable architectures for this specific domain.
3. Novel Ensemble Strategy: The WSB method introduces a unique strategy to bridge the gap between the best predictor and weaker models. By using the average of weaker, stable predictors to adjust the strongest predictor, the approach reduces bias and improves generalization, achieving lower error rates than individual models or traditional statistical methods.

The authors acknowledge that while the WSB approach improved performance, forecasting 12 months ahead remains challenging due to error accumulation in recursive strategies and unforeseen events (e.g., the 2023 consumption peak). The study concludes that while the proposed method is promising, future work should explore shorter forecast horizons and expand the methodology to geographically diverse institutions to further validate its applicability.