Evolutionary scheduling of university activities based on consumption forecasts to minimise electricity costs

TL;DR

The paper tackles reducing campus electricity costs through a predict-then-optimize framework that forecasts both demand and solar generation for November 2020 and optimizes class scheduling alongside battery use. It combines a LightGBM-based forecasting pipeline with feature-rich inputs (calendar, occupancy, trend/seasonality) and an evolutionary/base-two-stage optimization: (i) CMA-ES or GA to generate a base activity schedule, refined by local search, and (ii) a MIP solved in Gurobi to schedule batteries, all under precedence and room constraints. The key contributions include integrating multi-dimensional forecasts with large-scale optimization for campus demand response, demonstrating that battery-driven optimization yields significant cost reductions and that CMA-ES can outperform GA on large instances, with detailed performance and timing analyses. The results show the approach achieving near-top performance in a competition setting, offering practical insights into forecast accuracy, optimization strategies, and the value of separating activity and battery scheduling for large-scale problems.

Abstract

This paper presents a solution to a predict then optimise problem which goal is to reduce the electricity cost of a university campus. The proposed methodology combines a multi-dimensional time series forecast and a novel approach to large-scale optimization. Gradient-boosting method is applied to forecast both generation and consumption time-series of the Monash university campus for the month of November 2020. For the consumption forecasts we employ log transformation to model trend and stabilize variance. Additional seasonality and trend features are added to the model inputs when applicable. The forecasts obtained are used as the base load for the schedule optimisation of university activities and battery usage. The goal of the optimisation is to minimize the electricity cost consisting of the price of electricity and the peak electricity tariff both altered by the load from class activities and battery use as well as the penalty of not scheduling some optional activities. The schedule of the class activities is obtained through evolutionary optimisation using the covariance matrix adaptation evolution strategy and the genetic algorithm. This schedule is then improved through local search by testing possible times for each activity one-by-one. The battery schedule is formulated as a mixed-integer programming problem and solved by the Gurobi solver. This method obtains the second lowest cost when evaluated against 6 other methods presented at an IEEE competition that all used mixed-integer programming and the Gurobi solver to schedule both the activities and the battery use. The code and data used for the paper are publicly available.

Figures (5)

• Figure 1: Data flow of forecasting (in green) and schedule optimisation (in blue).
• Figure 2: Comparison of actual aggregated load vs the forecasted values from Monday 23 of November 2020 to the end of the scheduling time. In red is the forecasted load of the buildings minus the production of the solar panels. The blue line indicates the actual load.
• Figure 3: Visualisation of the optimisation process to find a base schedule with the GA and CMA-ES approach for different population sizes. Each solid line represents the summed best cost obtained for the 5 instances of each category (small in Fig. \ref{['fig:long_small']} and large in Fig. \ref{['fig:long_large']}) during consecutive optimisation processes. Once the evolutionary process reached its stopping criteria, a new evolutionary process for the same instance is started. The horizontal dashed lines are the costs of the schedules that obtained the lowest cost for each instance during all the runs and at the three phases of the scheduling process.
• Figure 4: Box-plots of the different methods used to improve the base schedule and the battery schedule. The box-plots for the improvement methods are generated on 40 values each, which are the costs of the improvement of the 8 best base solution found for each instance. The box plots for the battery schedule contain 80 values each, which are the cost obtained by adding the battery schedule to the 160 improved schedules from the 4 other box-plots.
• Figure 5: Representation of the best found improved schedule for the small_0 instance. In red is the forecasted load of the buildings minus the production of the solar panels. The load from the recurrent activities are represented in different colors for each activity. The orange vertical lines represent the start and the end of the working hours of the problem. The blue line is the price of electricity.