Haewon Byeon ^1* Azzah AlGhamdi ² Ismail Keshta ³ Mukesh Soni ⁴ Sultonali Mekhmonov ⁵ Gurpreet Singh ⁶

• ¹ Inje University, Gimhae, Republic of Korea
• ² Department of Computer Science, College of Computer Science and Information Technology, Imam Abdulrahman Bin Faisal University, Damam, Saudi Arabia
• ³ Department of Computer Science and Information Systems, College of Applied Sciences, University of Almaarefa, Dariyah, Riyadh, Saudi Arabia
• ⁴ Dr. D. Y. Patil Vidyapeeth, Pune, Dr. D. Y. Patil School of Science & Technology, Tathawade, Pune, Pune, India
• ⁵ Tashkent State Economic University, Tashkent, Uzbekistan
• ⁶ Chitkara University, Chandigarh, Punjab, India

Urban power load estimates are a critical component of smart grid scheduling and planning. Nevertheless, accurate forecasting is significantly obstructed by the issue of data imbalance. Conventional single-model methodologies fail to rectify this disparity, but current multi-model forecasting techniques partition the information into various groups according to power demand distribution and develop distinct models for each subset. While these methods somewhat alleviate data imbalance problems, they incur substantial model development expenses and may lead to the dissociation of common power distribution traits across various sample distributions. A model named DLUPLF, which denotes lightweight urban power load forecasting, is proposed as a remedy to these issues. This model enhances LSTM networks by integrating contrastive loss in short-term sampling with difference compensation. It additionally incorporates a communal feature extraction layer to diminish model development expenses. The difference compensating mechanism analyzes discrepancies between samples with diverse power load distributions and modifies the outcomes of the primary sequence prediction module accordingly. Regularization of model training using dynamic class-center contrastive learning loss is achieved by the loss resulting from contrastive sampling in the short term. To evaluate the model's effectiveness, experiments were conducted via parameter modification and comparison. The results indicate that the model accurately forecasts future power loads and model performs best when both hyperparameters HHH and CCC are set to 96 for predicting the next 10 days and 30 days. Therefore, based on the parameter tuning, we set both HHH and CCC to 96 in this study.