Scholarlink

<p>Heat waves, defined as prolonged periods of extremely high temperatures, seriously threaten human health, agriculture, ecosystems, and critical infrastructure. As global warming escalates, heat waves will become more frequent and severe, so it is important to have robust methodologies for forecasting. Typically, traditional weather prediction models suffer from non-stationarity and uncertainty in weather data, diminishing their reliability. This paper introduces a hybrid heat wave forecasting architecture called Empirical Model Decomposition (EMD) based Long Short Term Memory (LSTM) with Monte Carlo Quantile Regression (MCQR) and Federated Learning (FL-ELMQ) model, a federated learning framework that combines EMD, LSTM networks with MCQR to address the challenges presented. FL-ELMQ facilitates localized model training at each weather station, broadcasting only the model updates to a central hub, hence minimizing bandwidth and computational demands. To further enhance stability in the presence of heterogeneous and non-IID data across stations, the framework incorporates the FedProx approximation, which modifies the local training objective by adding a proximal term. This system addresses data uncertainties and manages non-stationary situations to enhance the accuracy of heat wave predictions. FL-ELMQ is evaluated using extensive data sets from IMD and ERA5, producing a mean absolute error (MAE) of 1.4471, a Mean Square Error (MSE) of 3.3862, a Root Mean Square Error (RMSE) of 1.8402, a Nash-Sutcliffe efficiency (NSE) of 0.9288, and a Correlation Coefficient (CC) of 0.9887. The results demonstrate that the combination of EMD and LSTM with MCQR approaches can successfully provide reliable heat wave forecasts under challenging climatic conditions.</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

FL-ELMQ: a hybrid federated learning framework to address non-stationarity and uncertainty in heat wave forecasting

  • 期刊论文

  • 发表日期：

    2025年8月2日

查看全文
  • Kaviya P,
  • Chitra P

摘要

Heat waves, defined as prolonged periods of extremely high temperatures, seriously threaten human health, agriculture, ecosystems, and critical infrastructure. As global warming escalates, heat waves will become more frequent and severe, so it is important to have robust methodologies for forecasting. Typically, traditional weather prediction models suffer from non-stationarity and uncertainty in weather data, diminishing their reliability. This paper introduces a hybrid heat wave forecasting architecture called Empirical Model Decomposition (EMD) based Long Short Term Memory (LSTM) with Monte Carlo Quantile Regression (MCQR) and Federated Learning (FL-ELMQ) model, a federated learning framework that combines EMD, LSTM networks with MCQR to address the challenges presented. FL-ELMQ facilitates localized model training at each weather station, broadcasting only the model updates to a central hub, hence minimizing bandwidth and computational demands. To further enhance stability in the presence of heterogeneous and non-IID data across stations, the framework incorporates the FedProx approximation, which modifies the local training objective by adding a proximal term. This system addresses data uncertainties and manages non-stationary situations to enhance the accuracy of heat wave predictions. FL-ELMQ is evaluated using extensive data sets from IMD and ERA5, producing a mean absolute error (MAE) of 1.4471, a Mean Square Error (MSE) of 3.3862, a Root Mean Square Error (RMSE) of 1.8402, a Nash-Sutcliffe efficiency (NSE) of 0.9288, and a Correlation Coefficient (CC) of 0.9887. The results demonstrate that the combination of EMD and LSTM with MCQR approaches can successfully provide reliable heat wave forecasts under challenging climatic conditions.