Research on Salt Pool Ion Concentration Prediction Model Based on LSTM and Attention Mechanism

The Lop Nur Salt Lake in Xinjiang is rich in underground brine resources and serves as an important potassium fertilizer production area in China. Potassium, an essential nutrient for plant growth, not only improves plant quality but also enhances their resistance to lodging, drought, cold, and lodging, playing a crucial role in agricultural production. In the process of ion collection in Lop Nur Salt Lake, the predicted ion concentration serves as a significant reference for guiding the industrial automation of salt mining machines. The predicted ion concentration is not only a part of automated mining but also a crucial step in the automated separation and purification of potassium salt crystals, as the ion concentration at the collection site directly affects the method of impurity removal. The predicted ion concentration can guide the automated production of potassium salt, and its accuracy directly impacts the efficiency of automated potassium salt production. To improve the accuracy of current salt pond ion concentration predictions and reduce prediction costs, this paper proposes a model integrating LSTM and attention mechanisms for time series prediction. We collected measurement data for various ion concentrations such as K<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">+</sup>, Na<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">+</sup>, $\text{SO}_4^{2 -}$, Cl<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−</sup>, and Mg<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2+</sup> using ion concentration measurement sensors, recording features such as collection time, location coordinates, salt field number, temperature, and air humidity. By training and testing the sample set using different models and evaluating and comparing the experimental results using various performance evaluation metrics, we found that the proposed model can effectively model long time series data, outperforming RNN (Recurrent Neural Network) and CNN (Convolutional Neural Network) in terms of prediction performance.