Improving Long-term Drought Forecasting with a novel Hybrid Deep Learning model based on Standardized Groundwater Level Index

Among other water resources, surface water, subsurface water, groundwater, and water supply are all adversely impacted by drought, a natural occurrence. As a type of hydrological drought, groundwater drought reflects both the unique features of the aquifer and human caused disturbances to the hydrological system. It is evident that human activity has both direct and indirect effects on the worsening of groundwater drought. Groundwater withdrawals are frequently used to meet water needs during hydrological and agricultural droughts because groundwater storage offers resilience. As a result, excessive groundwater extraction may make drought more severe. Quantitatively characterizing groundwater drought is extremely difficult due to the complex nature of groundwater flow systems and the difficulties in obtaining field observations pertaining to aquifers. By offering early warnings, long-term drought forecasting is essential to reducing drought risks.

Accurate long-term drought forecasting has long been of interest to researchers, but it is difficult because accuracy typically declines with forecasting period. This study's main goal is to present a novel hybrid deep learning model, Deep Feedforward Natural Networks (DFFNN), improved by War Strategy Optimization (WSO), for high accuracy long lead time drought forecasting. One of the vital aquifers in Odisha (Keonjhar district) was monitored for groundwater drought using the Standardized Groundwater Level Index (SGI), and forecasts were made for a range of lead times, including 1, 3, 6, 9, 12, and 24 months. For this study, monthly groundwater level data from 10 observation wells over a 25-year period (1996–2021) were collected. The observation wells were chosen based on their uniform distribution within the aquifer area and the completeness of their data records. The WSO algorithm was used to optimize important DFFNN parameters, such as the number of neurons and layers, learning rate, training function, and weight initialization. Two well known optimizers, Particle Swarm Optimization (PSO) and Grey Wolf Optimization, were used to validate the model's performance. 

Outcomes revealed that DFFNN-WSO model attained superior performance for SGI 24 (t + 12) with a coefficient of determination (r²) of 0.9847, Root Mean Square Error (RMSE) of 0.1035, willmott index of agreement (IoA) of 0.9812; for SGI 24 (t + 9) with r² = 0.8965, IoA = 0.8906 and RMSE = 0.1942; for SGI 12 (t + 6) with r² = 0.8473, IoA = 0.8352 and RMSE = 0.2315; for SGI 24 (t + 3) with r² = 0.7915, IoA = 0.7846 and RMSE = 0.2693; and for SGI 24 (t + 1) with r² = 0.7725, IoA = 0.7642 and RMSE = 0.3187 at the W5 station. Analysis of results indicated that DFFNN-WSO model outperformed all applied models consistently at all locations, and it considerably enhanced drought forecasting accurateness, with highest improvements for SGI 24 (t + 12) and moderate gains for SGI 24 (t + 1). The suggested model is a useful tool for real time drought monitoring and management since it offers precise and timely drought predictions, allowing for well-informed decision making to lessen the effects of drought.  

Keywords: Deep Feedforward Natural Networks (DFFNN); War Strategy Optimization; Standardized Groundwater Level Index (SGI); Water scarcity; Keonjhar