A COA-BP Model Incorporating Incremental Learning Mechanism and Its Linear Prediction

In the linear control of cantilever construction for long-span continuous rigid-frame bridge, existing prediction methods exhibit systemic deficiencies in both model construction and learning mechanisms. The traditional methods suffer from weak nonlinear fitting capabilities, while machine learning models are prone to local optima or insufficient generalization performance. In addition, most methods adopt a static modeling paradigm characterized by offline training and fixed parameters, which makes it difficult to dynamically adapt to the time-varying characteristics of structural responses and the accumulation of errors during construction. To overcome this problem, this study proposes a COA-BP model combining crayfish optimization algorithm (COA) and BP neural network, and innovatively introduces an incremental learning mechanism. Firstly, based on FEA NX, a refined solid finite element model was established. Considering the variability of key parameters such as concrete unit weight, elastic modulus and prestressed tension control stress, the Latin hypercube sampling was used to generate the input parameter combination, and the theoretical formwork elevation of each beam section was inversely calculated. The measured elevation was obtained after the completion of the site construction, and the difference between the two was used as the output target of the model. Then, the COA algorithm was used to optimize the initial weights and thresholds of the BP network, effectively improving the model’s convergence speed and global search capability. On this basis, a phased learning strategy was designed: segments No. 3, 4, and 5 were used for the static learning phase, where the model is initialized using the differences between measured and theoretical elevations. From block 6 onwards, the incremental learning phase begins, during which the model guides the adjustment of formwork placement elevations based on prediction results. After each construction phase, the elevation differences corresponding to newly measured data were incorporated into the training set, enabling a dynamic closed loop of “construction, learning, and optimization in parallel.” The proposed method is validated using an actual continuous rigid-frame bridge project. The results show that after the correction at segment No.6, the maximum error is reduced to −1.8 mm, and the prediction errors for subsequent beam segments converge continuously, with a smoothly declining prediction curve. This performance is significantly superior to that of traditional methods, demonstrating the effectiveness of the proposed COA-BP model in improving the accuracy and adaptability of linear prediction.