Abstract: To address the inefficiency of traditional AUVs using pre-programmed observation for harmful algal blooms and the insufficient adaptability of traditional prediction models to nonlinear concentration fields, a chlorophyll concentration prediction algorithm based on Bayesian Optimization Convolutional Neural Network (BO-CNN) was designed. The hybrid chlorophyll concentration field prediction model, based on convolutional neural networks and Bayesian optimization algorithms, trains the CNN on pre-sampled data, iteratively optimizes the hyperparameters of the CNN model using the Bayesian optimization algorithm, and finally input the optimized hyperparameters into the model to obtain the chlorophyll concentration prediction field at neighboring locations. By calculating the concentration gradient value between the current AUV position and the target position and combining it with the relative position to harmful algal bloom hotspots, different observation strategies are selected to dynamically guide the AUV’s movement direction, enabling it to autonomously and efficiently complete the search and observation tasks in high-concentration harmful algal bloom areas. Simulation tests using real harmful algal bloom chlorophyll concentration fields show that, compared to conventional algorithms, this algorithm reduces MAE by 30% and MSE by 60%, meeting the observation requirements for harmful algal bloom hotspot areas of different shapes.

Key words: Autonomous Underwater Vehicle, online path planning, adaptive sampling, hotspot area observation