Scholarlink

<p>Rainfall variability across the northeastern region (NER) of India exhibits pronounced spatial heterogeneity and strong seasonal dependence driven by complex atmospheric interactions. Using long-term monthly observations from 1981 to 2024, this study investigates the complex and scale-dependent variability of rainfall and its key atmospheric drivers. Traditional statistical and machine-learning approaches often fail to adequately capture these nonlinear dynamics across multiple temporal scales. To address this limitation, we propose a statistically integrated framework combining Granger causality (GC) and wavelet coherence (WTC) to identify and validate the influence of key atmospheric variables, namely relative humidity, minimum temperature, maximum temperature, wind speed, and wind direction, across five meteorological subdivisions. Relative humidity and minimum temperature consistently influence rainfall across all time scales (8-128 months), while maximum temperature and wind direction and wind speed show effects during the monsoon season. Building on these findings, we modified a novel WSG-Hybrid (Wavelet-Scale guided Variance Decomposition) integrated with an Autoregressive with Exogenous Inputs–Support Vector Regression (ARX-SVR) model to generate 24-month ahead forecasts. The model is evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Symmetric Mean Absolute Percentage Error (SMAPE), Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), Mean Absolute Scaled Error (MASE) and Percentage Bias(PBIAS). Results show that the WSG-Hybrid framework achieves the lowest RMSE, MAE, and MASE in four out of five subdivisions (ARP, ASML, GWB, SHWBS), with NSE values up to 0.874 and KGE up to 0.953, indicating superior predictive performance, while NBeatsX performs better in NMMT. Residual diagnostics using the Ljung–Box test confirm adequacy in four regions (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p&gt; 0.05\)</EquationSource> </InlineEquation>), and Diebold–Mariano (DM) tests establish statistically significant improvements at the 95% confidence level. Overall, the findings demonstrate that the WSG-Hybrid framework provides stable, accurate, and interpretable long-range rainfall forecasts suitable for climate-sensitive planning and resource management.</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

Wavelet-Scale guided Hybrid modelling for long-range rainfall forecasting in northeast India

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Kuldeep Goswami,
  • Junmoni Saikia

摘要

Rainfall variability across the northeastern region (NER) of India exhibits pronounced spatial heterogeneity and strong seasonal dependence driven by complex atmospheric interactions. Using long-term monthly observations from 1981 to 2024, this study investigates the complex and scale-dependent variability of rainfall and its key atmospheric drivers. Traditional statistical and machine-learning approaches often fail to adequately capture these nonlinear dynamics across multiple temporal scales. To address this limitation, we propose a statistically integrated framework combining Granger causality (GC) and wavelet coherence (WTC) to identify and validate the influence of key atmospheric variables, namely relative humidity, minimum temperature, maximum temperature, wind speed, and wind direction, across five meteorological subdivisions. Relative humidity and minimum temperature consistently influence rainfall across all time scales (8-128 months), while maximum temperature and wind direction and wind speed show effects during the monsoon season. Building on these findings, we modified a novel WSG-Hybrid (Wavelet-Scale guided Variance Decomposition) integrated with an Autoregressive with Exogenous Inputs–Support Vector Regression (ARX-SVR) model to generate 24-month ahead forecasts. The model is evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Symmetric Mean Absolute Percentage Error (SMAPE), Nash-Sutcliffe Efficiency (NSE), Kling-Gupta Efficiency (KGE), Mean Absolute Scaled Error (MASE) and Percentage Bias(PBIAS). Results show that the WSG-Hybrid framework achieves the lowest RMSE, MAE, and MASE in four out of five subdivisions (ARP, ASML, GWB, SHWBS), with NSE values up to 0.874 and KGE up to 0.953, indicating superior predictive performance, while NBeatsX performs better in NMMT. Residual diagnostics using the Ljung–Box test confirm adequacy in four regions ( \(p> 0.05\) ), and Diebold–Mariano (DM) tests establish statistically significant improvements at the 95% confidence level. Overall, the findings demonstrate that the WSG-Hybrid framework provides stable, accurate, and interpretable long-range rainfall forecasts suitable for climate-sensitive planning and resource management.