需水预测是区域水资源开发利用中的关键环节，鉴于江苏省需水影响因素涉及范围广、样本系列值少、需水波动幅度大等特点，提出了一种粒子群优化与正态区间估计联合优化的KS-PSO-BP耦合神经网络模型，模型使用主成分分析法与灰色关联分析法筛选需水影响因子，加入粒子群优化算法优化BP神经网络，通过引入K-S正态检验与正态区间估计，构建KS-PSO-BP神经网络需水预测模型，并对江苏省需水量进行模拟预测。结果表明：PSO-BP神经网络在预测结果相对误差上优于BP神经网络，对江苏省2021—2023年进行需水预测时，PSO-BP神经网络预测相对误差区间分别为-1.55%~1.14%、-1.65%~1.23%、-1.64%~1.18%，基于KS-PSO-BP耦合神经网络模型预测相对误差区间分别为-0.23%~0.09%、-0.28%~0.01%、-0.28%~0.02%，可见基于KS-PSO-BP耦合神经网络模型极大缩小了误差区间，模型预测结果更稳定、更准确，与江苏省实际用水状况更接近，该模型可作为江苏省需水预测的一种方法。

[Objective] Current research on water demand prediction primarily focuses on improving prediction accuracy， while studies on interval prediction of water demand remain limited. This study aims to further improve prediction accuracy， reduce error interval and reflect the actual regional water demand by incorporating the Kolmogorov-Smirnov （K-S） normal interval estimation into BP （back propagation） neural network model. [Methods] In view of a wide range of water demand influencing factors， limited sample series， and significant demand fluctuations in Jiangsu Province， we coupled the K-S normal interval estimation with particle swarm optimization （PSO） and BP neural network. Specific processes are as follows： the degree of influence and the number of key factors were preliminarily identified through principal component analysis （PCA）， followed by the calculation of grey relational grades （GRG） to determine the final water demand influencing factor index. This index was utilized as input for both BP and PSO-BP neural networks. The final model was then selected by comparing performance indicators， including relative error， Nash-Sutcliffe efficiency （NSE）， and coefficient of determination （R²）. Using multiple fitting iterations of the PSO-BP neural network， the sample size was expanded， and the K-S normality test and normal interval estimation were performed to further narrow the relative error intervals of the water demand prediction results. [Results] 1） Eight primary influencing factors of water demand were identified for Jiangsu Province： population， GDP， added value of secondary industry， added value of tertiary industry， per capita urban domestic water consumption， per capita rural domestic water consumption， irrigation water quota per mu （1 mu≈666.7 m²）， and water consumption per 10 000 yuan of industrial added value. 2） Both BP and PSO-BP neural networks exhibited strong performance in fitting historical water demand data， demonstrating their feasibility for future water demand prediction. Specifically， the PSO-BP neural network outperformed the standard BP network in the relative error of water demand prediction. The maximum relative errors for the training， validation， and testing samples were 2.23%， 0.88%， and 1.19%， respectively， with an average training error of 1.03%. The NSE and R² reached 0.98 and 0.99， respectively. 3） The selection of sample size significantly influenced the prediction results of water demand. With the increase of the sample size， the average value of samples was closer to the real value， and the overall prediction results of samples were more stable. The increase of sample size would reduce the benefit of improving the accuracy of the overall prediction results. 4） The K-S normality test and normal interval estimation stabilized the prediction results， substantially narrowed the error intervals， and better reflected the actual water demand. At a 95% confidence level， the relative error intervals of water demand prediction in 2021， 2022， and 2023 in Jiangsu Province were reduced to -0.23%-0.09%， -0.28%-0.01%， and -0.28%-0.02%， respectively. [Conclusion] The KS-PSO-BP neural network coupled model significantly reduces the error range and provides more stable and accurate prediction results that closely align with actual water demand in Jiangsu Province. The model serves as an effective method for regional water demand prediction and offers valuable guidance for future water resources planning in Jiangsu Province.