A watershed water quality prediction model based on attention mechanism and Bi-LSTM

Abstract

Accurate prediction of water quality is conducive to intelligent management and control of watershed ecology. Water quality data has time series characteristics, and although methods such as LSTM can capture sequence correlations in time series data, these methods do not consider the impact of bidirectional neighborhoods on the model, and they are not able to pay attention to the feature sequences to varying degrees. Aiming at this problem, this paper proposes a watershed water quality prediction model based on attention mechanism and bidirectional LSTM neural network (AT-BILSTM). The model mainly contains the Bi-LSTM layer and the temporal attention layer, and the attention mechanism is introduced after the bidirectional feature extraction of the water quality time series data to highlight the data series that have a key impact on the prediction results, and effectively use the sequence correlation of the watershed water quality data to improve the accuracy of the model. Finally, the actual datasets of four monitoring sites in the Lanzhou section of the Yellow River Basin in China were used to verify the effectiveness of the method. After comparing the reference model, the results show that the proposed model combines the bidirectional nonlinear mapping ability of Bi-LSTM and the feature-weighted characteristics of the attention mechanism. Taking FuHe Bridge as an example, compared with the original LSTM model, the RMSE and MAE of the model are reduced to 0.101 and 0.059, respectively, and the R2 is increased to 0.970, which has the best prediction performance in four sections, which can provide decision-making basis for comprehensive water quality management and pollutant control in the river basin.

DeePMD-kit v2: A software package for deep potential models

“…An attention-based descriptor , which is proposed in the pretrainable DPA-1 51 model, is given by where represents the embedding matrix after additional self-attention mechanism 119 and is defined by the full case in Eq. (12) .…”

“…The datasets we used included water, 9,61 copper (Cu), 107 high entropy alloys (HEAs), 51,171 OC2M subset in Open Catalyst 2020 (OC20), 115,116 Small-Molecule/Protein Interaction Chemical Energies (SPICEs), 104 and dipeptide subset in SPICE, 104 as shown in Table I and listed as follows: The water dataset contains of 140 000 configurations collected from path-integral ab initio MD simulations and classical ab initio MD simulations for liquid water and ice. Configurations were labeled using the hybrid version of Perdew–Burke–Ernzerhof (PBE0) 172 + Tkatchenko–Scheffler (TS) functional and projector augmented-wave (PAW) method.…”

“… The High Entropy Alloy (HEA) dataset comprises six elements: Ta, Nb, W, Mo, V, and Al. 51,171 These elements occupy a 2 × 2 × 2 BCC lattice consisting of 16 atoms in a random arrangement. The concentrations of Ta, Nb, W, Mo, and V encompass the entire composition space, while Al is considered an additive, with its maximum quantity being less than six.…”

“…Therefore, it is not advisable to utilize these models for production purposes, and it would be meaningless to compare them with the well-established models reported in other references, Refs. 9 , 51 , 107 , and 104 .…”

“…These systems include metallic materials, 55 non-metallic inorganic materials, 56–60 water, 61–71 organic systems, 10,72 solutions, 52,73–76 gas-phase systems, 77–80 macromolecular systems, 81,82 and interfaces. 83–87 Furthermore, the DeePMD-kit is capable of simulating systems containing almost all Periodic Table elements, 51 operating under a wide range of temperature and pressure, 88 and can handle drug-like molecules, 72,89 ions, 73,76 transition states, 75,77 and excited states. 90 As a result, the DeePMD-kit is a powerful and versatile tool that can be used to simulate a wide range of atomistic systems.…”