针对H-ADCP流量在线监测存在的特征流速选择困难、计算复杂度大及流量成果精度不高等难点,综合考虑特征数据降维、神经网络、支持向量机以及粒子群算法等人工智能算法,构建了基于特征自适应优选的H-ADCP流量在线监测模型,即FAO(Feature Adaptive Optimization)模型。选择受潮汐影响、水位流量关系复杂的罗湖水文站为研究对象,采用2019年和2023年实测流量资料,从各种算法的优缺点、降维数据的适用性以及模型稳定性等多个方面分析了FAO模型的适用性。结果表明:FAO模型具有较好的自学习能力,实测流量样本较少时,具有较好的流速特征学习能力和流量预测精度,检验期流量预测样本的均方根误差RMSE为6.06 m³/s、决定系数R²达到了0.93。提出的研究方法和模型可为流量在线监测研究提供新的研究思路和方法。

[Objective] To address the challenges faced by Horizontal Acoustic Doppler Current Profilers (H-ADCP) in online discharge monitoring applications—specifically, the difficulty in selecting index velocity (feature cells), the insufficient non-linear expressiveness of traditional calibration models, and the poor generalization ability and high computational complexity of existing machine learning models under complex hydrodynamic conditions such as tides and engineering regulations—this paper aims to develop a new H-ADCP online discharge monitoring model that can automatically optimize velocity features, integrate the advantages of multiple algorithms, and improve model accuracy. This model is designed to address the complex non-linear mapping problem between high-dimensional velocity data and cross-sectional discharge, thereby enhancing the accuracy, stability, and automation of discharge monitoring. [Methods] A Feature Adaptive Optimization (FAO) model for H-ADCP online discharge monitoring was developed. The technical framework of this model comprised three core components: (1) feature dimensionality reduction: Principal Component Analysis (PCA) was applied to conduct initial dimensionality reduction on the high-dimensional velocity data from up to 128 cells generated by the H-ADCP, reducing subsequent computational complexity while preserving the main velocity distribution characteristics. (2) Multi-model parallel mapping: five machine learning models—Backpropagation (BP) Neural Network, Elman Neural Network, Radial Basis Function (RBF) Network, Generalized Regression Neural Network (GRNN), and Support Vector Machine (SVM)—were constructed in parallel to establish the non-linear mapping relationship between the dimension-reduced feature velocities and the measured cross-sectional discharge. (3) Global optimization and adaptive selection: the Particle Swarm Optimization (PSO) algorithm was utilized as a global optimization engine, with the Root Mean Square Error (RMSE) as the fitness function, to search within the feature subspace and model space through iterative optimization and adaptively determine the optimal combination of velocity cells, the best machine learning model, and its corresponding parameters. To validate the model’s performance, the Luohu Hydrological Station, which is affected by both tides and backwater effects from confluence and has a complex hydrological regime, was selected as the study area. The model was calibrated and verified using measured H-ADCP velocity data and comparative discharge data from a moving-boat ADCP for the years 2019 and 2023. [Results] (1) The FAO model demonstrated superior performance: during the 2019 model verification period, the discharge predictions of the FAO model showed a high degree of agreement with the measured values, with a RMSE of 6.06 m³/s and a Coefficient of Determination (R²) reaching 0.93. This was significantly better than the traditional linear regression model and any single machine learning model. In simulating extreme discharges such as flood peaks, the FAO model also demonstrated a greater ability to capture them, with an annual maximum discharge error of 1.56%. (2) The feature optimization was effective: the model successfully and automatically selected an optimal combination of 11 feature cells ({5,9,12,15,17,19,21,24,26,28,35}) from 40 velocity measurement cells, eliminating invalid data affected by riverbanks and blind zones. The distribution pattern of the selected cells was highly consistent with hydraulic characteristics, demonstrating the physical interpretability of the model’s feature selection. (3) The model showed strong stability: when validated with data from the entire year of 2023, the FAO model performed stably, with an RMSE of 6.02 m³/s and an R² of 0.91, and effectively fitted the entire annual discharge process, especially for maximum and minimum values. [Conclusion] The proposed FAO model, by organically integrating PCA, multiple machine learning algorithms, and the PSO optimization algorithm, successfully addresses the key technical challenges in H-ADCP online discharge monitoring. The model exhibits powerful self-learning and self-adaptive capabilities, enabling it to automatically find the optimal velocity features and computational model based on data samples, while ensuring computational accuracy and significantly reducing data processing complexity. The application case under complex hydrological conditions demonstrates that the FAO model has high accuracy, good stability, and strong adaptability, providing an efficient and intelligent solution for H-ADCP online discharge monitoring.