Aquaculture water quality has a very important impact on the output and income of aquaculture. Predicting the water quality state in advance can improve the governance level and thus reduce the loss of aquaculture. With daily aquaculture water quality of Penaeus vannamei (known as white-leg shrimp) as the object of study, and the temperature (T), pH, dissolved oxygen (DO), salinity, oxidation reduction potential (ORP), NO2--N and NH4+-N  selected as water quality data, a water quality prediction model based on local Bi-LSTM(CovBiLSTM) and state transformation constraint is proposed. First, Bi-LSTM is used to receive the input historical water quality data sequence information, then the convolution function and max pooling are used to mine the relationship between the output data of different units in Bi-LSTM for integration of water quality data in different historical periods, and finally Softmax classifier is used to predict the water quality state and the state transformation constraint is used to improve the accuracy of prediction. The effectiveness of local Bi-LSTM and state transformation constraint in water quality prediction is demonstrated by comparative experiments. Compared with the prediction method based on LSTM, the classification accuracy and recall rate of evaluation indexes are improved by 5% and 4% respectively. The results show that the water quality prediction algorithm based on CovBiLSTM network model can predict the water quality state of Penaeus vannamei more accurately than that based on other models.