[Objective] Existing short-term power load forecasting models struggle to simultaneously learn the inherent differences in electricity consumption patterns across diverse scenarios and effectively capture long-range temporal dependencies

so their forecasting accuracy is significantly restricted in high-volatility situations such as extreme weather events

peak load periods

or sudden changes in user behavior. To address this limitation

this paper proposes a novel hybrid forecasting framework named “similar-day clustering-temporal convolutional network(TCN)-Informer”. [Methods] First

the fuzzy C-means(FCM) clustering algorithm is employed to categorize historical load samples into three distinct similar-day clusters(working days

weekends

and holidays) based on historical load data and calendar features(e.g.

weekday/weekend flags

holiday labels). This preprocessing step effectively mitigates cross-pattern noise interference caused by heterogeneous consumption behaviors

laying a solid foundation for subsequent feature extraction. Second

for each clustered dataset

a causal dilated TCN is utilized to extract local temporal features and capture short-to medium-term load fluctuations

as its dilated convolution structure enables efficient expansion of the receptive field without increasing computational complexity while avoiding future information leakage. Subsequently

the Informer model equipped with a probabilistic sparse attention mechanism is introduced to capture long-range dependencies in load time series

which significantly reduces the computational burden compared to traditional attention mechanisms and enhances the model's ability to learn long-term temporal correlations. Moreover

exogenous variables closely related to power load

including meteorological factors(temperature

humidity

solar irradiation) and real-time electricity prices

are integrated into the forecasting framework to improve the comprehensiveness of input information. To optimize the model's hyperparameters

Bayesian optimization is adopted to efficiently search the global hyperparameter space

outperforming traditional grid search in both search efficiency and optimization performance. [Results] Experimental validations for 24-hour load forecasting are conducted using 30-minute resolution data from the AEMO-NSW power market

covering a one-year period with over 17 000 samples. The results demonstrate that the proposed model achieves remarkable performance: its root mean square error(RMSE) and mean absolute error(MAE) reduce by 16.59% and 14.65% compared with those of the classic LSTM model

and by 12.33% and 11.60% in contrast to the advanced TimesNet model

with the mean absolute percentage error(MAPE) as low as 3.02%. Particularly in high-volatility scenarios

the model maintains superior stability and accuracy. [Conclusion] These findings fully verify the prominent advantages of the proposed method in forecasting precision and robustness

which can provide reliable decision-making support for power grid dispatching

electricity spot trading

and energy resource optimization in practical power systems.