Scholarlink

<p>The primary challenge in forecasting power consumption and PV power generation is their dependence on several inherent variables and sometimes unpredictable complex factors. As the inclusion of renewable energy resources increases in modern utility grids, accurate forecasting becomes inevitable in managing an efficient, sustainable, stable, and safe smart grid. Traditional methods struggle with the irregular nature of the influential factors on electrical load as well as the fluctuation of weather conditions which have a significant impact on forecasting accuracy of PV power generation as well as residential power consumption forecasts. This study tackles the challenge of achieving accurate forecasting and the integration of the influence of weather conditions as an aspect often neglected in previous work to investigate its impact on prediction accuracy. A novel hybrid LSTM-CNN-Transformer model -based architecture is proposed in this paper for forecasting power consumption and PV power generation in a hybrid microgrid, providing an enhanced alternative to traditional predictive models presented in earlier studies. The proposed model achieves high prediction accuracy by assigning attention to the important components within the time series data and effectively capturing complex temporal patterns through the transformer mechanism. Furthermore, the inclusion of the LSTM-CNN layers to the proposed model enhances the model’s capability to extract both temporal and spatial features from the time series. The suggested model is evaluated using real-world datasets for PV power generation and power consumption collected from a suburban substation in southern UK. The simulation results demonstrate that the proposed model achieves significantly lower forecasting errors over extended forecasting horizons; 6&#xa0;h, 12&#xa0;h and 24&#xa0;h compared to the other competitive models, based on RMSE and MAE metrics. Specifically, the proposed model outperforms the GRU baseline model, reducing RMSE by 30.6% for 6-hour power-consumption forecasts without weather data, and achieves an additional 8.3% RMSE improvement over the CNN–LSTM autoencoder model for 12-hour PV power-generation predictions. Moreover, incorporating weather data enhances PV forecasting accuracy, while it is not the case for short term power consumption forecasting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Advanced Transformer Model-Based LSTM-CNN for Power Consumption and PV Power Generation Forecasting in Modern Microgrid

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Mohamed Sayed Ibrahim,
  • Sawsan Morkos Gharghory,
  • Hanan Ahmed Kamal

摘要

The primary challenge in forecasting power consumption and PV power generation is their dependence on several inherent variables and sometimes unpredictable complex factors. As the inclusion of renewable energy resources increases in modern utility grids, accurate forecasting becomes inevitable in managing an efficient, sustainable, stable, and safe smart grid. Traditional methods struggle with the irregular nature of the influential factors on electrical load as well as the fluctuation of weather conditions which have a significant impact on forecasting accuracy of PV power generation as well as residential power consumption forecasts. This study tackles the challenge of achieving accurate forecasting and the integration of the influence of weather conditions as an aspect often neglected in previous work to investigate its impact on prediction accuracy. A novel hybrid LSTM-CNN-Transformer model -based architecture is proposed in this paper for forecasting power consumption and PV power generation in a hybrid microgrid, providing an enhanced alternative to traditional predictive models presented in earlier studies. The proposed model achieves high prediction accuracy by assigning attention to the important components within the time series data and effectively capturing complex temporal patterns through the transformer mechanism. Furthermore, the inclusion of the LSTM-CNN layers to the proposed model enhances the model’s capability to extract both temporal and spatial features from the time series. The suggested model is evaluated using real-world datasets for PV power generation and power consumption collected from a suburban substation in southern UK. The simulation results demonstrate that the proposed model achieves significantly lower forecasting errors over extended forecasting horizons; 6 h, 12 h and 24 h compared to the other competitive models, based on RMSE and MAE metrics. Specifically, the proposed model outperforms the GRU baseline model, reducing RMSE by 30.6% for 6-hour power-consumption forecasts without weather data, and achieves an additional 8.3% RMSE improvement over the CNN–LSTM autoencoder model for 12-hour PV power-generation predictions. Moreover, incorporating weather data enhances PV forecasting accuracy, while it is not the case for short term power consumption forecasting.