Accurately adjusting the feeding intake is often required in pond aquaculture of Micropterus salmoides. In this study, a dual-model progressive feeding strategy was proposed to accurately forecast the postprandial feeding status of the subsequent round using a time series model. The decrement mechanism was first triggered. Subsequently, a classification model was utilized to determine the postprandial feeding status of the current round, enabling the termination of the decrement mechanism. The experimental system consisted of three pond culture tanks stocked with Micropterus salmoides, with initial body masses of (161.47±11.02) , (204.36±17.09) , and (220.25±22.78) g, and the stocking densities of 224, 301, and 294 individuals, respectively. Feeding experiments were conducted using a multi-round feeding protocol within a single session. Audio data was collected during feeding using the digital hydrophone, while video data was obtained using a camera. Meanwhile, multimodal data including light intensity, water temperature, body mass, per-round feeding rate, and population abundance were synchronously recorded for per-round feeding. The stabilization of the cumulative feeding energy curve was adopted to determine feeding termination. A dynamic adaptive threshold segmentation was employed to accurately identify the feeding audio within the collected audio. Principal component analysis was conducted on the feeding features to extract from the feeding audio. Feature selection was employed to identify sensitive features to feeding status, which served as the inputs for model training and classification. Four models 1D convolutional neural network, Long short-term memory, gated recurrent unit, and transformer were improved to construct the feeding prediction models with time series. The optimal model was selected to predict the postprandial feeding status of the subsequent round, serving as the first-stage model of feeding strategy. Eight models k-Nearest Neighbors, decision tree, support vector machine, random forests, adaBoost, gradient boosting decision trees, extreme gradient boosting, and light gradient boosting machine (LightGBM) were employed to construct feeding status classification models. The best classification performance was selected as the feeding determination model in the second stage. Results showed that the sensitive features included overall features, light intensity, water temperature, body mass, and per-round feeding rate. A classification between the original and sensitive features demonstrated the effectiveness of the sensitive features, according to gradient boosting decision trees and random forests models. Transformer-regression-classification (transformer-RC) model outperformed the rest of the models across 1 to 3-time windows, with the accuracy from 0.93 to 0.94. The Transformer-RC model effectively predicted the postprandial feeding status of the subsequent round. All five ensemble models achieved strong classification, which outperformed the three base models. Among them, the LightGBM model achieved an accuracy of 0.98 for inputs to the classification model of feeding status in the second stage of the feeding strategy. Both the Transformer-RC and LightGBM models shared high prediction and classification after validation, with average values of four evaluation metrics exceeding 0.89 under both feeding states, indicating strong generalization. This finding can provide a strong reference to develop intelligent feeding.