Data-Driven Optimization of Energy Consumption Management in Smart Grids

Abstract

Smart grids today are increasingly confronted with the difficulty of balancing rising demand in the face of renewable integration and carbon reduction goals, while maintaining reliability and cost effectiveness. Traditional forecasting and optimization techniques face challenges with heterogeneous data and dynamic control. This paper presents a data driven optimization framework to address these issues utilizing the OpenEI Smart Grid dataset. Data was preprocessed using regression imputation and data augmentation to fill missing values, normalization to account for differences in scale between variables, and feature engineering (time of day, day of week, and seasonal indicators) to enhance predictive capability. Subsequently, three models were applied to the data: a deep neural network (DNN) load forecasting model, a reinforcement learning (RL) agent which dynamically controlled the grid, and a RL – DNN flexible model which leveraged accuracy in prediction with adaptable optimization. Hyperparameters were selected via grid and random search search resulting in an RL accuracy of 89.4%, DNN accuracy of 91.2% and hybrid model accuracy of 92.5%.

The simulations of residential, commercial, and industrial scenarios showed noticeable enhancements. The energy efficiency rate improved in the three sectors by 10.9%, 12.9%, and 12.4%, respectively. The cost savings were 14.4% ($219), 14.0% ($823), and 13.8% ($1,556), respectively. Carbon emissions showed reductions of 15.8%, 14.8%, and 15.9%, respectively. The hybrid model demonstrated a superior performance than the one based on one of the approaches by an overall grid optimization improvement of 12%, peak demand reduction of 18%, and a 14% reduction in the period of renewable surpluses compared to grid reliance on conventional energy, while maintaining reasonable computational efficiency. These results support the conclusion that utilizing forecasting and real-time optimization capabilities together through a hybrid RL–DNN model can provide measurable energy efficiency, cost, and emissions savings. For international energy policy makers, the findings promote the acceleration of smart grid intelligence through interoperable infrastructure, privacy-preserving data environments, and pilot projects to scale up these models into operational surroundings.