Machine Learning for Reducing Energy Costs in Urban Grids

Thesis Abstract

This study addresses the escalating energy costs in urban electricity networks by developing and validating a machine learning (ML) framework that optimizes energy consumption, pricing signals, and distributed energy resources to enhance grid efficiency and reduce expenditures for both utilities and end-users. The core aim is to design an adaptive, data-driven system that leverages urban smart meter data, weather information, and grid topology to forecast demand, detect A/C and industrial load anomalies, and implement dynamic adjustment strategies without compromising reliability. Specific objectives include (1) designing an ML-based demand forecasting model with hourly granularity achieving at least 95% forecast accuracy over a 7-day horizon; (2) developing a reinforcement learning (RL) policy for real-time demand response and distributed energy resource (DER) dispatch that reduces peak demand by 8–12% in simulation and pilot deployments; (3) evaluating the impact of price-aware control on household and commercial energy expenditure; (4) assessing grid reliability resilience under ML-driven control using probabilistic risk metrics; and (5) providing a decision-support toolkit for utility operators that integrates interpretability methods to explain model recommendations. The methodology adopts a mixed-methods, action-research design conducted in three phases over 24 months. The population comprises urban residential, commercial, and industrial customers (n=15,000) connected to a metropolitan distribution network and a sample of 100 grid operators. Data collection instruments include smart meter time-series (15-minute intervals for 12 months), SCADA and substation telemetry, feeder topology data, weather station observations, tariffs, and outage logs, complemented by semi-structured interviews with 20 utility engineers to capture operational constraints. Instrument validation employs content validity checks with domain experts and test-retest reliability for time-series preprocessing, achieving a reported Cronbach’s alpha of 0.88 for reliability of survey components. The ML components integrate a hybrid forecasting model combining gradient boosting and long short-term memory (LSTM) networks to capture nonlinear temporal patterns and seasonality, and a deep Q-learning RL agent embedded within a model-predictive control (MPC) framework to determine DER dispatch and demand response actions. Model interpretability is addressed through SHAP analysis and feature attribution to ensure transparency for operators. The analysis plan includes (i) out-of-sample evaluation of forecast accuracy using RMSE and MAE, (ii) simulation-based stress testing of the RL policy against historical extreme events, and (iii) econometric evaluation of price-response effects using panel data regression with fixed effects and instrumental variables to control for endogeneity. Hypothesis testing centers on whether ML-driven control significantly reduces peak demand and total energy expenditure while maintaining grid reliability above a pre-set reliability threshold (SAIDI/SAIFI). Expected findings indicate that the hybrid ML forecasting model achieves a mean absolute percentage error (MAPE) under 4% across peak and off-peak periods, and the RL-enabled MPC reduces peak demand by 9–13% in simulated scenarios, with corresponding reductions in customer bills by an average of 7–11% depending on tariff structure. Anticipated improvements in grid reliability metrics, particularly SAIDI and SAIFI within regulatory targets, are projected due to more precise scheduling of DERs and demand shaping. The study also expects to identify critical drivers of energy savings, such as weather-normalized load sensitivity, occupancy-based usage patterns, and tariff design, with SHAP analyses revealing actionable insights for policy and operational decisions. The contribution to knowledge includes (a) an integrative ML-DER control framework for urban grids, (b) empirical evidence on the effectiveness of price-aware, data-driven demand management in reducing energy costs, and (c) a transferable methodology and toolkit for utility operators to deploy interpretable ML-based optimization in live networks. The main conclusion posits that carefully architected ML models, when coupled with operator-centric interpretability and robust validation, can meaningfully decrease urban energy costs while preserving reliability. Recommendations emphasize scalable deployment strategies, stakeholder engagement for tariff adaptation, continuous model monitoring, cybersecurity safeguards for ML systems, and further research into multi-objective optimization under evolving renewable penetration.