Research On Prediction Method Of Shrimp Feeding Based On Machine Learning

In recent years,More and more people buy aquatic products,Therefore,breeding efficiency needs to be improved and reducing aquaculture costs to meet the growing demand.Among them,bait is the most critical shrimp pond culture can manage the cost of regulation.When the feed exceeds the feed intake of prawns,the residual feed falls into the bottom of the water,which not only wastes the feed,increases the harmful bacteria or toxic substances in the aquaculture water,but also requires more funds to maintain water quality and increases the maintenance cost of aquaculture.On the contrary,if the feeding amount is less than the feed intake of shrimp,shrimp will lack sufficient nutrients to grow,the growth rate becomes particularly slow,and the intake of shrimp is not uniform,shrimp body length and weight will have a significant gap,uneven growth.At present,shrimp pond culture in China is still in the stage of determining the feeding amount based on the experience of shrimp farmers and then putting it into shrimp ponds.There are certain errors in the feeding amount determined by experience.In this paper,a prediction model of shrimp feeding quantity based on LSTM is constructed.The input variables of LSTM are extracted by CNN and the deep internal relations of these input variables are excavated.The self-attention mechanism is used to focus on the timing data information of important time nodes and learn.The CNN-LSTM-ATTN shrimp feeding quantity prediction model is established.Finally,the CNN-LSTM-ATTN prediction model is optimized by IPSO,which reduces the influence of empirical setting parameters on the model and make the CNN-LSTM-ATTN shrimp bait prediction model become more stabilty and accuracy.This paper mainly includes:(1)Factors affecting shrimp feeding were analyzed,including external factors of aquaculture environment,internal factors of shrimp itself and aquaculture management factors.In this paper,the relationship between factors such as aquaculture environmental factors,aquaculture management factors and shrimp growth cycle and the required feeding amount was considered.The average weight,number,water temperature and dissolved oxygen of shrimp were selected as the input variables of the prediction model,and the total feeding amount required for aquaculture was selected as the output variable(2)For the shrimp feeding amount data are a set of nonlinear time series data,the LSTM shrimp feeding amount prediction model is constructed,and the parameter settings of the prediction model are determined through the control variable experiment.Finally,the constructed LSTM prediction model was used to predict the feeding amount of the test set samples.The model can roughly fit the actual feeding amount of shrimp,but there are certain errors.(3)Aiming at the problem of insufficient ability of LSTM prediction model to mine learning input variables,a CNN-LSTM shrimp bait prediction model is constructed by using CNN to extract feature from LSTM input data,mining and learning the internal relationship between input variables.Compared with the results and performance evaluation indexes of CNN-LSTM prediction model and LSTM prediction model,CNNLSTM prediction model has higher prediction accuracy,but the error between the predicted value and the actual value is still obvious.(4)The CNN-LSTM prediction model did not consider the significant changes in the amount of bait required for shrimp in different growth periods.The self-attention mechanism can be used to learn the importance of time series of different time nodes and assign different weights,so as to reduce the prediction error caused by the difference of shrimp feeding amount in different growth stages.The CNN-LSTM-ATTN shrimp feeding amount prediction model is constructed.Compared with the results of other prediction models,it is found that the prediction error of the model at the final sample point is smaller,but the prediction error of the intermediate sample is larger.The accuracy of the model cannot be directly determined.By comparing the performance evaluation indexes of these prediction models,the RMSE,MAE and MAPE of CNN-LSTM-ATTN prediction model are 0.816,0.681 and 0.018,respectively,which are the minimum values.So,CNN-LSTM-ATTN prediction model has the best prediction accuracy and the best model stability.(5)In order to reduce the error caused by human experience parameter setting on the prediction model,IPSO algorithm was proposed to automatically optimize the parameters of CNN-LSTM-ATTN prediction model,and the prediction model of IPSO-CNN-LSTMATTN shrimp feeding amount was constructed.The results of IPSO-CNN-LSTM-ATTN prediction model are closer to the actual feeding amount.Compared with the feeding amount calculated by feeding rate,it was found that the IPSO-CNN-LSTM-ATTN prediction model had higher prediction accuracy,with an average error of 1.2 %.Finally,the prediction model is used to predict the feeding amount of shrimp pond.Compared with the feeding amount determined by shrimp farmers ’ experience,the prediction model is reasonable,but it also needs to consider the operation of bottom change and adding bacteria in shrimp pond culture.