|
Peak load of electricity demand forecast using machine learning and electricity generation planning |
|---|---|
| รหัสดีโอไอ | |
| Title | Peak load of electricity demand forecast using machine learning and electricity generation planning |
| Creator | May Thazin Phuu Wai |
| Contributor | Aussadavut Dumrongsiri, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Classification, Short-term load forecasting, Machine learning, Random forest classification, Calendar-based segmentation |
| Abstract | Short-term load forecasting plays a critical role in power system planning, operational scheduling, and economic dispatch. However, forecasting accuracy often deteriorates under irregular calendar conditions, such as weekends, public holidays, and bridging holidays, when load behavior deviates from typical daily patterns. This study proposes a two-stage hybrid forecasting framework that integrates calendar-aware classification with machine learning regression to improve day-ahead load prediction across diverse operating conditions. The methodology incorporates Random Forest (RF) classification to segment historical data using Month of Year (MoY), Day of Week (DoW), holiday, and bridging-holiday indicators, followed by RF regression to predict the 48 half-hourly loads for Thailand and the 24-hourly loads for France from 2019 to 2021. A linear interpolation mechanism is introduced to address insufficient samples in rare calendar categories.Experimental results demonstrate that the proposed RF-RF framework consistently outperforms baseline methods, including Multiple Linear Regression (MLR), Support Vector Regression (SVR), Everyday classification, and Rule-based classification across both countries. For Thailand, the hybrid model achieves the lowest average MAPE of 4.03% and RMSE of 4.47%, effectively capturing nonlinear seasonal and calendar-driven variations. For France, characterized by strong winter heating demand, the proposed method also yields superior performance, with MAPE 3.01% and RMSE 4.49%, confirming its generalizability across different climatic and load-profile regimes. The improvements are most pronounced on holidays and bridging holidays, where traditional models typically suffer from instability due to irregular consumption patterns. Overall, this research demonstrates that integrating calendar-based segmentation with ensemble learning enhances pattern recognition, model robustness, and prediction accuracy. The proposed framework offers a scalable, interpretable solution for system operators seeking reliable short-term forecasting across diverse climatic contexts and complex calendar effects. |
