Abstract: To enhance the interpretability of power grid load forecasting at the sector level and improve the precision of weather attribution, this paper proposes a three-layer analytical framework comprising "daily aggregate STL decomposition – intraday actual proportion reconstruction – sector-specific meteorological regression." Based on 96-point load data from 37 national economic sectors in the Guangzhou power grid during 2021–2025, three consolidation principles are established—parent-child deduplication, homonym normalization, and volume threshold control—to integrate raw entries into 11 decomposition units. Holiday and production schedule effects are first stripped out via pre-regression; hierarchical STL is then employed to extract trend and multi-period seasonal components; and actual intraday load proportions are introduced to correct baseline load reconstruction deviations. Finally, a dual-track Ridge/OLS meteorological regression model is established for individual sectors. The results indicate that the temperature elasticity of the electric power and heat production and supply industry reaches 243.82 MW/℃, representing the primary stress source during system summer peaks; the computer, communication and other electronic equipment manufacturing sector exhibits the strongest model explanatory power (R2 = 0.242); and the transportation sector"s load is dominated by operational schedules, showing extremely low meteorological sensitivity. The proposed method effectively separates baseline load from weather-sensitive load, providing a data foundation for grid dispatching departments to formulate sector-specific and temperature-interval-based forecast correction strategies.