Scholarlink

The need for accurate weather prediction models is on the rise, considering their potential to make impactful decisions in various sectors of society. Numerical Weather Prediction (NWP) models have excelled with accurate predictions when data was limited. In the Big Data era, the availability of a large amount of diverse weather data has been a challenge to the forecasting power of the traditional models. Deep learning entered the picture as an alternative that can learn accurate patterns from complex datasets. In this paper, we implement a model that predicts short-term temperature based on a Stacked Long Short-Term Memory (LSTM) on historical Radar data collected from Cochin University of Science & Technology. An Explainable AI using SHapley Additive exPlanations (SHAP) is implemented on the model to determine the effect of different features on the predicted values. The proposed model is compared with a single layer vanilla LSTM and the performance is evaluated based on Root Mean Square Error (RMSE), Mean Absolute Error (MAE), R2 score and Mean Absolute Percentage Error (MAPE). The results showed that the model outperforms the vanilla LSTM and performs on par with state-of-the-art methods. The SHAP results identify features that have an upper hand in the final predictions.

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

Explainable AI Insights into a Time Series Weather Prediction Model Using Stacked LSTM

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • T. H. Sunu Fathima,
  • Binsu C. Kovoor

摘要

The need for accurate weather prediction models is on the rise, considering their potential to make impactful decisions in various sectors of society. Numerical Weather Prediction (NWP) models have excelled with accurate predictions when data was limited. In the Big Data era, the availability of a large amount of diverse weather data has been a challenge to the forecasting power of the traditional models. Deep learning entered the picture as an alternative that can learn accurate patterns from complex datasets. In this paper, we implement a model that predicts short-term temperature based on a Stacked Long Short-Term Memory (LSTM) on historical Radar data collected from Cochin University of Science & Technology. An Explainable AI using SHapley Additive exPlanations (SHAP) is implemented on the model to determine the effect of different features on the predicted values. The proposed model is compared with a single layer vanilla LSTM and the performance is evaluated based on Root Mean Square Error (RMSE), Mean Absolute Error (MAE), R2 score and Mean Absolute Percentage Error (MAPE). The results showed that the model outperforms the vanilla LSTM and performs on par with state-of-the-art methods. The SHAP results identify features that have an upper hand in the final predictions.