An atmospheric visibility prediction method based on neural network model fusion

Abstract

The application discloses an atmospheric visibility prediction method based on neural network model fusion, relates to the cross technical field of artificial intelligence, atmospheric science and meteorology, and comprises the following steps: acquiring meteorological station and pollution monitoring station data, constructing an hour-level time series dataset after time-space matching and gridding processing; based on an XGBoost and Bi-LSTM neural network model, a fusion model is constructed, the hour-level time series dataset is used for training, a prediction model is constructed, and the prediction model is used for inputting historical 10h observation data by using a sliding window to obtain a prediction result of future 1h visibility. The application solves the deficiencies of traditional models and existing neural network models in visibility prediction, and provides a scientific basis for fog and haze early warning, traffic control and regional joint prevention and control.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of artificial intelligence, atmospheric science, and meteorology, and more specifically, to an atmospheric visibility prediction method based on neural network model fusion. Background Technology

[0002] Visibility prediction is a core research area in meteorological monitoring and ecological environment early warning, playing a vital supporting role in transportation, environmental governance, and disaster prevention and mitigation. Currently, visibility prediction technologies are mainly divided into three categories: traditional physical models, statistical models, and emerging machine learning models. Each method has certain technical limitations, making it difficult to meet the demands for high-precision, real-time, and lightweight visibility prediction.

[0003] Traditional forecasting methods primarily rely on physical models such as atmospheric chemical transport models and statistical models such as multiple linear regression. Physical models simulate and predict based on atmospheric physical and chemical mechanisms, but they suffer from high computational complexity and strong model parameter constraints, making it impossible to update forecast results in real time and resulting in poor engineering applicability. Statistical models have simple structures and fast computation speeds, but they have weak fitting capabilities for complex nonlinear correlation characteristics in the atmospheric environment. When faced with special meteorological scenarios such as extreme pollution and dense fog, the prediction accuracy drops significantly, and the generalization ability is insufficient.

[0004] To overcome the shortcomings of traditional models, machine learning and deep learning methods have been widely applied in the field of visibility prediction in recent years. Existing research has controlled the mean absolute error (MAE) of 24-hour visibility prediction to within 0.8 km by constructing risk-weighted neural networks, optimizing the training strategy of backpropagation neural networks, and building a single LSTM time-series prediction model. At the same time, the industry has also developed improved solutions such as combining numerical models with neural networks and LSTM-Adaboost hybrid models, which have effectively improved the ability to model the temporal variation patterns of visibility and significantly improved the prediction performance compared with traditional methods.

[0005] Although existing deep learning and hybrid prediction models have significant advantages in modeling time series features, they still have several key technical shortcomings in practical applications, which seriously restrict the overall performance and practical application effectiveness of visibility prediction models. First, the spatial feature extraction capability is insufficient. Single time-series models such as LSTM can only mine the temporal dimension features of data and cannot effectively capture the regional spatial dependence of atmospheric visibility, resulting in limited accuracy in spatiotemporal coupling modeling. Second, hyperparameter tuning is inefficient. Existing model parameter optimization relies heavily on grid search or manual experience selection, which is time-consuming and labor-intensive, and it is difficult to guarantee obtaining the globally optimal parameter combination. Third, multi-source data fusion is insufficient. Meteorological data and atmospheric pollution monitoring data have insufficient spatiotemporal alignment accuracy, and the model feature selection lacks interpretability, with redundant features easily interfering with prediction results. Fourth, model deployment costs are high. High-precision models such as Transformer have problems with large memory consumption and high computational resource consumption, making it difficult to achieve lightweight and embedded scenario deployment. Fifth, the prediction capability for extreme events is weak. Existing models show a significant decrease in feature fitting and generalization capabilities under extreme weather scenarios such as pollution peaks and dense fog, failing to meet the accurate prediction requirements under extreme conditions. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide an atmospheric visibility prediction method based on neural network model fusion, which addresses the shortcomings of traditional models and existing neural network models in visibility prediction and provides a scientific basis for haze warning, traffic control, and regional joint prevention and control.

[0007] To achieve the above technical objectives, this application provides an atmospheric visibility prediction method based on neural network model fusion, comprising the following steps: Data from meteorological stations and pollution monitoring stations were acquired, and after spatiotemporal matching and gridding, hourly time-series datasets were constructed. Based on XGBoost and Bi-LSTM neural network models, a fusion model is constructed and trained using hourly time-series datasets to build a prediction model. This model is used to obtain the visibility prediction results for the next hour by using a sliding window to input historical 10-hour observation data.

[0008] Preferably, during the spatiotemporal matching and gridding process, radiosonde data, ground meteorological station data, atmospheric pollutant data, and atmospheric boundary layer data are acquired, and missing value filling is performed. Then, the time is aligned to an hourly scale, and the space is interpolated to a unified grid. The ground meteorological station data includes temperature, humidity, air pressure, and wind speed; the atmospheric pollutant data includes PM2.5, PM10, SO2 concentration, NO2 concentration, and O3 concentration.

[0009] Preferably, when acquiring hourly time-series datasets, after normalizing the processed observation data, individual cases of low visibility events occurring 1 hour before and during the event are selected from the observation data based on the minimum visibility value, and a database of low visibility time cases of the process length series is compiled to obtain hourly time-series datasets.

[0010] Preferably, when acquiring hourly time-series datasets, the observed data are also subjected to Gaussian smoothing and noise reduction processing.

[0011] Preferably, when constructing the fusion model, the XGBoost model is used to obtain the visibility prediction results, and the contribution of features to visibility is quantified by the SHAP value to screen key features. The key features include PM2.5, humidity, wind speed, PM10, and temperature.

[0012] Preferably, when using the XGBoost model, the XGBoost model is configured as follows: tree depth is limited to 6, initial learning rate is 0.1, row sampling ratio is 0.8, column sampling ratio is 0.8, and a random seed is saved for reproducing the results.

[0013] Preferably, when constructing the fusion model, a deep learning prediction model containing an input layer, hidden layer, fully connected layer and output layer is constructed based on a bidirectional Bi-LSTM network, and then combined with the XGBoost model to form a fusion model.

[0014] Preferably, when constructing a deep learning prediction model, the parameter gradient is calculated using the chain rule to obtain the LSTM backpropagation gradient.

[0015] Preferably, when constructing the prediction model, the time-series sample data obtained from the XGBoost model is divided into a training set and a validation set through a cross-validation mechanism. The model training is completed by automatically tuning based on Bayesian optimization, using RMSE as the loss function and Adam as the optimizer.

[0016] Preferably, when constructing the prediction model, the learning rate is adjusted by a piecewise constant decay.

[0017] The present invention discloses the following technical effects: (1) High prediction accuracy: The RMSE reached 793.3813 on the test set, and the loss value was reduced to 0.0088.

[0018] (2) Strong robustness: It performs stably in extreme pollution and meteorological events.

[0019] (3) Good interpretability: The contribution of each feature is clearly defined by the SHAP value, which improves the transparency of the model.

[0020] (4) Excellent real-time performance: Supports hourly forecasting to meet business deployment requirements. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the method described in this invention.

[0023] Figure 2 This is a diagram showing the training results of visibility prediction as described in this invention.

[0024] Figure 3 This is the loss curve diagram described in this invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] Example 1: As Figure 1 As shown, this invention provides an atmospheric visibility prediction method based on XGBoost and Bi-LSTM neural network models. The method of this invention integrates meteorological characteristic factors, atmospheric pollutant concentrations, atmospheric boundary layer conditions and atmospheric circulation data to construct a visibility forecasting model that combines a bidirectional long short-term memory network (Bi-LSTM) with a multi-head self-attention mechanism. It is suitable for real-time visibility prediction in complex terrain and areas with concentrated pollution.

[0027] Step S100: Multi-source data fusion and preprocessing.

[0028] In step S100 of one embodiment, data from weather stations and pollution monitoring stations are acquired, and after spatiotemporal matching and gridding processing, an hourly time-series dataset is constructed. D = {time step, number of features, number of samples} For example, in step S100, step S101 acquires radiosonde data, observation data from ground meteorological stations, atmospheric pollutants, and atmospheric boundary layer, performs missing value imputation, aligns the time to an hourly scale, and spatially interpolates to a unified grid; the observation data from the ground meteorological stations includes temperature, humidity, air pressure, and wind speed; the atmospheric pollutant observation data includes PM2.5. 2.5 PM 10 The concentrations of SO2, NO2, and O3 are also measured, and the observed data need to be processed by Gaussian smoothing and noise reduction.

[0029] For example, in step S100, step S102 normalizes the observation data processed in step S101.

[0030] For example, in step S100, step S103 filters out cases of low visibility events from the observation data based on the minimum visibility value, including cases that occurred 1 hour before the event and those that occurred during the event, and organizes a database of low visibility time cases with a long duration sequence.

[0031] Step S200 Feature selection and optimization.

[0032] In step S200 of one embodiment, the visibility prediction result is obtained by using the XGBoost model, and the contribution of features to visibility is quantified by the SHAP value to screen key features, construct an optimized feature subset, and obtain time series sample data. For example, the XGBoost model is configured as follows: tree depth is limited to 6, initial learning rate is 0.1, row sampling ratio is 0.8, column sampling ratio is 0.8, and a random seed is saved for reproducing the results.

[0033] For example, in step S200, step S201 predicts the 10-minute average visibility using the XGBoost model and obtains the top five key features by filtering through SHAP values: PM2.5, humidity, wind speed, PM10, and temperature.

[0034] For example, in step S200, step S202 constructs a sliding window and obtains the visibility prediction result for the next 1 hour using historical 10-hour observation data.

[0035] Step S300: Deep learning model construction.

[0036] In step S300 of one embodiment, a deep learning prediction model comprising an input layer, a hidden layer, a fully connected layer, and an output layer is constructed based on a bidirectional Bi-LSTM network.

[0037] For example, in step S300, step S301 LSTM forward propagation settings include: (1) The LSTM forward propagation gating mechanism is as follows: In the formula, It's an input gate. It is the Gate of Oblivion. It's an output gate. It is the sigmoid function. It is the previous hidden state. , , These are the bias terms for each door. , , It connects to the current input. The weight matrix between each gate (i,f,o), , , It is the hidden state from the previous moment. The corresponding weight matrix for each gate, This is the input for the current time step; (2) The candidate memory units and memory update mechanism are as follows: In the formula, It is the result of memory cell state prediction. It connects to the current input. To candidate memory units The weight matrix, It connects to the hidden state of the previous moment. To candidate memory units The weight matrix; It is the current state of the memory cell; It is the state of the previous memory unit; It is a bias term; This represents element-wise multiplication; (3) Hidden output settings: In the formula, It is currently in a hidden state.

[0038] For example, in step S300, S302 LSTM backpropagation gradient settings: Calculate the parameter gradient using the chain rule: In the formula, Represents the loss function. This represents the weight parameter matrix in the LSTM model. For the gradient of the parameters, This represents the total number of time steps in the input sequence; The update of the weight matrix follows the Adam optimizer rules: in, It is a dynamically adjusted learning rate; k is the kth iteration, and k+1 is the next iteration.

[0039] For example, in step S300, the loss value is recorded in step S303: In each iteration, the loss on the mini-batch data is calculated as follows: In the formula, This represents the average loss over a small batch. Let represent the true visibility value of the i-th sample. This represents the visibility value predicted by the model for the i-th sample. For example, in step S300, the learning rate scheduling is set in step S304: The learning rate is adjusted by piecewise constant decay: In the formula, This represents the learning rate at the t-th training epoch. This is the initial learning rate, with a value of 0.0005; It is the attenuation factor, with a value of 0.5; It is the decay period, with a value of 10, decaying once every 10 epochs.

[0040] Step S400: Hyperparameter optimization and training.

[0041] In step S400 of one embodiment, cross-validation is used to divide the time-series sample data of step S200 into a training set and a validation set; the model training is completed by automatic tuning based on Bayesian optimization, using RMSE as the loss function and Adam as the optimizer.

[0042] Step S500, based on the optimized deep learning model, constructs a sliding window to input historical 10-hour observation data and obtains the prediction result of visibility for the next hour.

[0043] Example 2 aims to construct a visibility prediction method that integrates multi-source data, possesses strong generalization ability and high accuracy, and addresses the shortcomings of traditional models and existing neural network models in regional visibility prediction, providing a scientific basis for haze early warning, traffic control, and regional joint prevention and control. Figure 1 As shown, taking a region as an example, the atmospheric visibility prediction method based on XGBoost and Bi-LSTM neural network models of the present invention includes the following steps: Step S100 involves the collection, quality control, and preprocessing of various types of data.

[0044] For example, historical reanalysis data includes radiosonde data, surface meteorological station data (including temperature, humidity, air pressure, and wind speed), and air pollutant (PM2.5) data for the winter six months (October to February of the following year) from 2013 to 2023 provided by the meteorological bureau. 2.5 PM 10 Data collected included atmospheric boundary layer observations (SO2, NO2, O3, etc.). Fog observation data included daily fog records and visibility data for the region. Spatiotemporal matching and gridding were performed; Gaussian smoothing was applied for noise reduction to construct an hourly time-series dataset; a sliding window was used to construct input samples in the format [time step, number of features, number of samples].

[0045] In step S100 of one implementation, step S101 performs quality control on radiosonde data, ground station meteorological data, and atmospheric pollutant data. For single missing values, time series values ​​are used to fill the gaps; for long time series missing values, spatial nearest neighbor interpolation is used; other abnormal or erroneous values ​​are directly removed. Abnormal and null values ​​are removed; time alignment is done on an hourly scale; spatial interpolation is performed to a uniform grid; and Gaussian smoothing (window L=5, σ=1) is used for noise reduction.

[0046] In step S100 of one embodiment, the data normalization method in step S102 adopts a discrete normalization method: performing a linear transformation on the data to map the data values ​​to the range [0,1].

[0047] In step S100 of one embodiment, step S103 performs case screening on historical meteorological boundary layer observation data, atmospheric pollutant observation data, and meteorological observation data. Based on the minimum visibility value, cases that occurred 1 hour before and during the occurrence of low visibility events are selected. The cases include key meteorological elements, six elements of atmospheric pollution, and atmospheric boundary layer data, and are compiled into a long-sequence low visibility event case library.

[0048] Step S200 uses the XGBoost regression model to predict the 10-minute average visibility and analyzes the influence direction and importance of each feature on the prediction results through SHAP value analysis. That is, key features are selected through feature importance analysis (using XGBoost+SHAP) to provide a basis for improving the efficiency and accuracy of subsequent LSTM model visibility predictions. The top 5 key features selected are: PM 2.5 Humidity, wind speed, PM 10 Temperature; Construct a sliding window input: historical 10-hour data to predict visibility for the next 1 hour.

[0049] In step S200 of one implementation, step S201 configures the XGboost model as follows: tree depth is limited to 6, initial learning rate is 0.1, row sampling ratio is 0.8, column sampling ratio is 0.8, and a random seed is saved to facilitate the reproduction of results.

[0050] In step S200 of one embodiment, step S202 uses the XGBoost+SHAP feature selection → LSTM time series modeling process, which can take into account both feature interpretability and sequence prediction capability.

[0051] Step S300: Model Construction and Training.

[0052] In one embodiment, step S300 is based on a bidirectional LSTM network to construct a prediction model that includes an input layer (multivariate time series), a hidden layer (bidirectional LSTM units capture the temporal dependencies before and after the input), a fully connected layer, and an output layer (ReLU activation function) to avoid the gradient vanishing problem.

[0053] In step S300 of one implementation, step S301 designs the mathematical framework for LSTM forward propagation. The core of LSTM is the gating mechanism, and its mathematical expression is as follows: (1) Input gate, forget gate, output gate in: It's an input gate. It is the Gate of Oblivion. It's an output gate. It is the sigmoid function. It is the previous hidden state. It is a bias term. and It is a weight matrix. This is the input for the current time step.

[0054] (2) Candidate memory units and memory update In the formula, It is the result of memory cell state prediction. It is the current state of the memory cell; It is the state of the previous memory unit; It is a bias term; This indicates element-wise multiplication.

[0055] (3) Output in hidden state In the formula, It is currently in a hidden state.

[0056] In step S300 of one embodiment, step S302 obtains the LSTM backpropagation gradient.

[0057] For example, the parameter gradient can be calculated using the chain rule: ; The update of weight W follows the Adam optimizer rules: in, It is a dynamically adjusted learning rate.

[0058] In step S300 of one embodiment, step S303 records the mathematical mapping of the loss value.

[0059] For example, in each iteration, the loss on the mini-batch data is calculated as follows: .

[0060] The update rule for the global array `losses` is: when the k-th iteration is completed... .

[0061] In step S300 of one embodiment, S304 discusses the mathematical effects of learning rate scheduling.

[0062] For example, the learning rate is adjusted by a piecewise constant decay: In the formula, This represents the learning rate at the t-th training epoch. This is the initial learning rate, with a value of 0.0005; It is the attenuation factor, with a value of 0.5; It is the decay period, with a value of 10, decaying once every 10 epochs.

[0063] For example, a mathematical extension of bidirectional LSTM: the hidden state of bidirectional LSTM consists of a concatenation of forward and backward LSTMs. in: It is the hidden state of the forward LSTM. It is the hidden state of the backward LSTM.

[0064] Step S400: Hyperparameter optimization and training.

[0065] In step S400 of one embodiment, cross-validation is used to divide the time-series sample data of step S200 into a training set and a validation set; the model training is completed by automatic tuning based on Bayesian optimization, using RMSE as the loss function and Adam as the optimizer.

[0066] Step S500, based on the optimized deep learning model, constructs a sliding window to input historical 10-hour observation data and obtains the prediction result of visibility for the next hour.

[0067] Example 3: Taking the visibility prediction effect diagram of a single fusion site as an example, specifically, the feature column data of the past 10 hours is used to predict the target column data of the next hour.

[0068] For example, 5-fold cross-validation; early stopping mechanism (patience=10); loss function: RMSE+MAE.

[0069] For example, experimental results: Training loss: The final loss value is 0.0088, and the convergence curve is smooth (see...). Figure 3 ); Test performance: RMSE=793.3813, the predicted value and the actual value trend are consistent (see Figure 2 ); SHAP analysis: PM2.5 and humidity are the most significant negative factors, while wind speed is a positive factor; Extreme event validation: In the dense fog event in January 2019, the model prediction error was reduced by about 20% compared to the traditional LSTM.

[0070] For example, the feasibility and beneficial effects analysis are as follows: Feasibility: This invention is based on mature machine learning and deep learning frameworks (such as XGBoost, TensorFlow / PyTorch) and multi-source observation data. The model structure is reasonable and the training process is stable. Beneficial effects: By using XGBoost for pre-feature screening, it provides a basis for improving the efficiency and accuracy of subsequent LSTM model visibility prediction; by calling SHAP to calculate the contribution of various meteorological elements and pollutant parameters to visibility, it provides interpretable feature contribution analysis and scientific basis for decision-making; and it provides a lightweight and scalable solution for regional visibility prediction systems.

[0071] In summary, this invention integrates regional meteorological and pollution monitoring data, constructs hourly multi-source datasets using spatial interpolation techniques, and employs Gaussian smoothing algorithms to suppress temporal random fluctuations. Based on this, a spatiotemporal distribution characteristic map of pollutant concentrations in the study area is constructed, revealing the distribution patterns and evolution characteristics of pollutants under different weather conditions.

[0072] This invention employs a combined strategy of grid search and Bayesian optimization to determine the optimal combination of initial learning rate, batch size, number of LSTM units, and Dropout rate, and dynamically adjusts the learning rate to improve tuning efficiency.

[0073] This invention integrates multi-source data based on data from regional meteorological observation stations and pollutant monitoring stations using spatial interpolation techniques to construct a comprehensive dataset with a time resolution of 1 hour. Then, an XGBoboost regression model is used to predict the 10-minute average visibility, and the influence direction and importance of each feature on the prediction results are analyzed through SHAP values, improving the interpretability of feature selection.

[0074] This invention avoids using complex structures with excessive GPU memory consumption (such as Transformer) in its model design, opting instead for a lightweight LSTM + attention hybrid architecture. Future research could combine federated learning and edge computing to achieve low-power deployment of distributed visibility prediction systems.

[0075] In this invention, the evaluation of feature importance not only considers the global contribution but also analyzes the changes in feature weights under different seasons and weather conditions, thereby improving model adaptability. Model training employs K-fold cross-validation, with particular attention paid to prediction stability under different seasons and weather conditions.

[0076] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0078] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An atmospheric visibility prediction method based on neural network model fusion, characterized in that, Includes the following steps: Data from meteorological stations and pollution monitoring stations were acquired, and after spatiotemporal matching and gridding, hourly time-series datasets were constructed. Based on the XGBoost and Bi-LSTM neural network models, a fusion model is constructed. The model is trained using the hourly time series dataset to build a prediction model. This model is used to obtain the prediction result of visibility in the next hour by using a sliding window to input historical 10-hour observation data.

2. The atmospheric visibility prediction method based on neural network model fusion according to claim 1, characterized in that: During spatiotemporal matching and gridding, radiosonde data, ground meteorological station data, atmospheric pollutant data, and atmospheric boundary layer data are acquired. After missing value imputation, the time is aligned to an hourly scale, and the spatial data is interpolated to a unified grid. Ground meteorological station data includes temperature, humidity, air pressure, and wind speed; atmospheric pollutant data includes PM2.

5. 2.5 PM 10 SO2 concentration, NO2 concentration, O3 concentration.

3. The atmospheric visibility prediction method based on neural network model fusion according to claim 2, characterized in that: When acquiring hourly time-series datasets, after normalizing the processed observation data, individual cases of low visibility events occurring 1 hour before and during the event are selected from the observation data based on the minimum visibility value, and a database of low visibility time cases of the process length series is compiled to obtain the hourly time-series dataset.

4. The atmospheric visibility prediction method based on neural network model fusion according to claim 2, characterized in that: When acquiring hourly time-series datasets, Gaussian smoothing and noise reduction processing is also performed on the observed data.

5. The atmospheric visibility prediction method based on neural network model fusion according to claim 1, characterized in that: When constructing the fusion model, the XGBoost model is used to obtain visibility prediction results, and the contribution of features to visibility is quantified by the SHAP value to screen key features. These key features include: PM 2.5 Humidity, wind speed, PM 10 ,temperature.

6. The atmospheric visibility prediction method based on neural network model fusion according to claim 1, characterized in that: When using the XGBoost model, the XGBoost model is configured as follows: tree depth is limited to 6, initial learning rate is 0.1, row sampling ratio is 0.8, column sampling ratio is 0.8, and a random seed is saved for reproducing the results.

7. The atmospheric visibility prediction method based on neural network model fusion according to claim 1, characterized in that: When constructing the fusion model, a deep learning prediction model containing an input layer, hidden layer, fully connected layer and output layer is constructed based on a bidirectional Bi-LSTM network, and then combined with the XGBoost model to form the fusion model.

8. The atmospheric visibility prediction method based on neural network model fusion according to claim 7, characterized in that: When building a deep learning prediction model, the parameter gradient is calculated using the chain rule to obtain the LSTM backpropagation gradient.

9. The atmospheric visibility prediction method based on neural network model fusion according to claim 1, characterized in that: When building the prediction model, the time series sample data obtained from the XGBoost model will be divided into training and validation sets through cross-validation. The model training will be completed automatically by using RMSE as the loss function and Adam as the optimizer, based on Bayesian optimization.

10. The atmospheric visibility prediction method based on neural network model fusion according to claim 1, characterized in that: When building the prediction model, the learning rate is adjusted by a piecewise constant decay.

Images