In power system planning and operation, load forecasting is an important task as it helps ensure a reliable and efficient electricity supply. For effective operation of the smart grid, load forecasting is also an important thing to keep balancing dispatch of power, load management, and load shifting. In this regard, this paper aims to propose an accurate load forecasting based on implementing and integrating different load forecasting models using standalone machine learning and ensemble machine learning models, particularly for segregated real-world load data. In the given context, machine learning models namely, k-nearest neighbor, random forest, decision tree, and voting ensemble regression, are used in this study. The time series load data for this research work was acquired from a real-world load database namely, Pecan Street Dataport. For performance evaluation, two statistical error matrices are used, i.e., mean absolute error (MAE) and mean squared error (MSE). For simulation purposes, Python along with different machine-learning libraries was employed. Moreover, for numerical data analysis and visualization, this research work utilizes different packages like NumPy, pandas, and matplotlib. The empirical study presents the comparative performance analysis of machine learning models for load forecasting utilizing low sampling load data, both at aggregated as well as at segregated levels. Standalone and ensemble machine learning algorithms yield very good forecasting results, and this research has revealed that machine learning models trained on segregated data exhibit superior performance compared to those trained on aggregated data. On segregated data, the proposed voting- based ensemble machine learning algorithm outperforms all the other models with MAE 0.05708, followed by k-nearest neighbors (with MAE 0.05879), random forest (with MAE 0.07069), and decision tree (with MAE 0.07361).

In the face of escalating energy demand, this research proposes a demand-side management (DSM) strategy that focuses on appliance-level load shifting in residential environments. The proposed approach utilizes detailed energy consumption forecasts that are generated by ensemble machine learning models, which predict usage at both whole-household and individual appliance levels. This granular forecasting enables the development of customized load-shifting schedules for controllable devices. These schedules are optimized using a metaheuristic genetic algorithm (GA) with the objectives of minimizing consumer energy costs and reducing peak demand. The iterative nature of GA allows for continuous fine-tuning, thereby adapting to dynamic energy market conditions. The implemented DSM technique yields significant results, successfully reducing the daily energy consumption cost for shiftable appliances. Overall, the proposed system decreases the per-day consumer electricity cost from 237 cents (without DSM) to 208 cents (with DSM), achieving a 12.23% cost saving. Furthermore, it effectively mitigates peak demand, reducing it from 3.4 kW to 1.2 kW, which represents a substantial 64.7% reduction. These promising outcomes demonstrate the potential for substantial consumer savings while concurrently enhancing the overall efficiency and reliability of the power grid.