TL;DR
This paper presents a robust "predict + optimize" framework for managing smart building energy. By combining ensemble machine learning for load forecasting, deep residual networks for solar prediction, and a two-stage mixed-integer linear programming (MILP) approach for scheduling, the authors developed a system that minimizes electricity costs and grid dependence even when forecasts aren't perfect.
What Problem Does This Paper Address?
As we move toward "smart cities," managing the energy interaction between buildings, local solar generation, and battery storage becomes a massive logistical headache. The authors tackle the IEEE-CIS Technical Challenge, which asks: How do you schedule a month's worth of activities (like university lectures) and battery cycles to minimize costs while dealing with the inherent uncertainty of weather and human behavior?
The challenge is twofold:
- Forecasting: Predicting "baseload" (how much power the building uses naturally) and solar generation with high granularity (15-minute intervals).
- Optimization: Scheduling thousands of activities across multiple buildings. This is computationally "heavy" because the search space for scheduling grows exponentially with the number of activities.
Methods
The authors broke the problem into a clear pipeline:
1. Data Pre-processing
They used a rolling window approach to clean "noisy" historical data, replacing invalid points with moving averages or annual averages. They noted specific quirks—like one building having discrete power steps (1, 2, 3, 4 kW)—and tailored their cleaning to those patterns.
2. Load Forecasting (The "Demand" Side)
For building load, they used a Voting Regressor ensemble combining Random Forest (RF) and Gradient Boosting (GB). For buildings with complex seasonal cycles, they applied STL (Seasonal-Trend decomposition using Loess) to separate the data into trend, seasonality, and "noise" before training.
3. Solar Forecasting (The "Supply" Side)
Solar data is highly repetitive but weather-dependent. The authors used a Refined Motif (RM) discovery method to find the most common daily patterns. These patterns, along with weather data, were fed into a ResNet (Residual Network). They found that ResNet's deeper architecture allowed for better "grokking" (generalization through over-parameterization) compared to standard 1D-CNNs.
4. Two-Stage Optimization
To make the scheduling tractable, they split the problem:
- Sub-problem 1: Solve a Linear Programming (LP) relaxation to find a lower bound for the peak power limit.
- Sub-problem 2: Use that limit to solve the full MILP, scheduling activities and batteries to "fill the valleys" of the power demand curve, effectively flattening the load profile.
Datasets
The study uses the IEEE-CIS 3rd Technical Challenge dataset, which includes:
- Historical load data for six different buildings.
- Solar PV production data associated with those buildings.
- Wholesale electricity prices (AEMO).
- Exogenous variables like weather, occupancy, and calendar information.
Key Insights and Ideas
- The "Valley-Filling" Strategy: By focusing on minimizing peak power and filling in the "valleys" of demand with scheduled activities and battery charging, the system becomes naturally robust. Even if a forecast is slightly off, the flat power profile prevents expensive price spikes.
- ResNet for Time Series: While CNNs are common for time series, the authors found that ResNets (typically used in image recognition) were faster to train and more accurate for solar forecasting because they could handle deeper layers without vanishing gradient issues.
- Tractability vs. Perfection: The authors realized that scheduling non-recurring activities added massive complexity for a relatively small financial gain (~$16,000). They chose to focus only on recurring activities to ensure the optimization could actually finish running in a reasonable timeframe.
Key Results
- Accuracy: Their forecasting model achieved a Mean Absolute Scaled Error (MASE) of 0.67 at the end of the competition's first phase.
- Model Improvement: Moving from a simple Random Forest to their tuned ensemble and ResNet approach improved the MASE from 0.82 to 0.70.
- Computation Time: By using the two-stage optimization approach with the Gurobi solver, they achieved solution times of roughly 15 to 30 minutes per instance, making it practical for real-world monthly planning.
Limitations and Future Work
The authors noted that while deep learning (ResNet) worked well for solar, it was too computationally expensive to apply to building load features within the competition's constraints. Additionally, the current model ignores non-recurring activities to maintain speed.
Future work could involve more sophisticated "end-to-end" models where the forecasting loss function is directly tied to the optimization cost, rather than treating them as two separate steps.
Summary updated