Multi-pollutant concentration prediction method and system based on space-time attention and gating reset cycle network
By adopting a method based on spatiotemporal attention and gated reset recurrent networks, the problem of insufficient prediction accuracy and stability in multi-pollutant concentration prediction is solved. This method achieves high-precision and stable multi-step forward-looking prediction, and improves the interpretability and multi-scale representation capability of the model, making it suitable for urban environmental monitoring platforms.
Patent Information
- Application Number
- CN202511770842.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-02-10
AI Technical Summary
Existing technologies for predicting multi-pollutant concentrations suffer from limitations in prediction accuracy and stability, limited ability to fuse multi-source data, and insufficient model interpretability. In particular, they are insufficient in characterizing long-range time dependence and multi-scale spatial heterogeneity, leading to the accumulation of multi-step forward prediction errors and poor robustness.
We adopt a method based on spatiotemporal attention and gated reset recurrent network. Through the joint modeling framework of residual convolutional coding, dual spatiotemporal attention and attention-enhanced GRU, we fuse traffic and land use features with multi-scale concentric buffers, introduce dynamic reset gate to suppress error propagation, and use Smooth L1 loss and Adam optimizer for training.
It significantly improves the accuracy and stability of single-step and multi-step forward-looking predictions, enhances multi-scale characterization capabilities, improves the robustness and interpretability of the model, and facilitates integration and deployment with urban environmental monitoring platforms.
Smart Images
Figure CN121503815A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of environmental monitoring and air pollution prediction, and specifically discloses a multi-pollutant concentration prediction method and system based on space-time attention and a gated reset recurrent network. BACKGROUND
[0002] Air pollution is an important environmental risk affecting public health and urban management. The existing methods often face the following problems when dealing with multi-source heterogeneous data (meteorology, traffic, land use, etc.) and complex spatio-temporal dependence (long-short coupling, local / regional multi-scale difference): First, the long-range temporal dependence and multi-scale spatial heterogeneity are not well characterized, which limits the prediction accuracy and stability.
[0003] Second, the multi-source data fusion capability is limited, and the interactive influence of traffic flow-land use-meteorology cannot be considered; multi-step forward (3 / 6 / 12 / 24h) prediction is prone to error accumulation and drift, and the robustness is poor.
[0004] Finally, the model lacks explainability, which is not conducive to policy making and real-time management.
[0005] Therefore, there is an urgent need for a prediction method that can simultaneously extract residual spatial features, allocate spatio-temporal attention weights, suppress multi-step error propagation, and have explainability. SUMMARY
[0006] In view of the defects and problems of the existing multi-pollutant concentration prediction method, such as limited prediction accuracy and stability, limited fusion capability, and insufficient model explainability, the application provides a multi-pollutant concentration prediction method and system based on space-time attention and a gated reset recurrent network.
[0007] The application provides a multi-pollutant concentration prediction system based on space-time attention and a gated reset recurrent network, comprising: A data acquisition and alignment module for acquiring and aligning multi-source features of pollutants, meteorology, traffic and land use; A residual convolutional encoding module for multi-layer residual convolutional feature extraction of the aligned multi-source features; A space-time attention module for adaptively weighting and fusing the extracted feature key areas and key time periods to obtain space-time weighted features, and convolving the space-time weighted features to obtain a mixed feature map; An attention-enhanced gated recurrent module for inputting the mixed feature map into a recurrent unit according to time steps to generate a hidden state sequence; A prediction and regularization module for outputting a time step or multi-step forward prediction result according to the hidden state sequence and implementing Dropout; The training and optimization module is used for end-to-end training with Smooth L1 loss function, Adam optimizer and cosine annealing learning rate scheduling. The results output module is used to output the prediction results of each pollutant and make them available for subsequent business systems to use.
[0008] The aforementioned multi-pollutant concentration prediction system based on spatiotemporal attention and gated reset recurrent networks includes a spatial-temporal attention module comprising a spatial attention submodule and a temporal attention submodule. The spatial attention submodule focuses on the extracted features. Max pooling and average pooling are applied along the time dimension, and after concatenation along the channel dimension, a spatial weight map is generated by 3×3 convolution and sigmoid activation. Then, the features are weighted based on the spatial weight map to generate spatial attention features. The temporal attention submodule maps the features to query Q, key K, and value V, calculates the attention weights, and outputs the temporal attention features.
[0009] This invention also provides a method for predicting the concentration of multiple pollutants based on spatiotemporal attention and gating reset recurrent networks, comprising the following steps: S1. Obtain the pollutant concentration sequence, meteorological elements, traffic flow indicators and static land use attribute characteristics of the target area at hourly time resolution, and align the features with spatial location at all time steps. S2. The obtained aligned features are extracted using a residual convolution editor to obtain a high-level semantic representation. ; S3, Based on high-level semantic representation The spatiotemporal weighted features are obtained by fusing key spatial regions and key time segments. Then, the input consists of residual blocks composed of multiple 3×3 convolutions, normalization, and activation, and the output is a hybrid feature map; S4. Input the hybrid feature map into the attention-enhanced gated recurrent module at time steps to generate a hidden state sequence; S5. Input the hidden state sequence into the prediction and regularization module outputs the time-step or multi-step forward prediction results and implement Dropout. S6. Using Smooth L1 loss as the training objective, the Adam optimizer is employed in conjunction with cosine annealing learning rate scheduling for training optimization.
[0010] The multi-pollutant concentration prediction method based on spatiotemporal attention and gating reset recurrent networks described above, in step S1, includes pollutants such as CO, NO, NO2, O3, BC, and PM2.5. 2.5 PM 10The meteorological elements include air pressure, dew point, air temperature, humidity, precipitation, visibility, wind speed, and wind direction.
[0011] The above-mentioned multi-pollutant concentration prediction method based on space-time attention and gated reset recurrent network, the spatial representation of the traffic flow characteristics in S1 comprises: constructing a multi-scale concentric buffer zone of 0.1-5 km around each monitoring site, with a step length of 0.1 km, respectively extracting and aggregating the overall level and spatial difference features of the traffic flow, and aligning them one by one with the hour-level time series input according to the time stamp.
[0012] The above-mentioned multi-pollutant concentration prediction method based on space-time attention and gated reset recurrent network, step S3 comprises: (1) Feature Max-pooling and average-pooling are applied along the time dimension,
[0013] , , , After concatenation in the channel dimension, a 3x3 convolution and a Sigmoid activation are performed to generate a spatial weight map Then, the features are weighted to generate spatial attention features , , In the formula: represents element-wise multiplication; (2) Map the feature to query Q, key K, and value V, calculate the attention weight, and weight the output time attention feature with the attention weight , , ; In the formula: d is a scaling factor; (3) Fuse and to obtain space-time weighted features , and input the space-time weighted features into a residual block composed of multiple 3x3 convolutions, normalizations, and activations to output a mixed feature map.
[0014] The above-mentioned multi-pollutant concentration prediction method based on space-time attention and gated reset recurrent network, in step S4, the dynamic reset gate in the attention enhanced gated recurrent module injects the attention weight as a gating bias or scaling term into the GRU reset gate.
[0015] The multi-pollutant concentration prediction system based on the spatio-temporal attention and the gated reset recurrent network, in step S5, for the time step prediction, the prediction and regularization module inputs the hidden state sequence into a fully connected layer to obtain a one-step prediction of the next hour, and applies Dropout in the hidden or output layer to alleviate overfitting.
[0016] The multi-pollutant concentration prediction system based on the spatio-temporal attention and the gated reset recurrent network, in step S5, for the multi-step prediction, the prediction and regularization module adopts a stacked fully connected layer to output 3h, 6h, 12h, and 24h prediction values in parallel in the same forward process; or generates multi-look-ahead results in sequence through a recursive or time shift mechanism.
[0017] The application also provides a readable storage medium storing program instructions, and a computer program is stored thereon, which is executed by a processor to implement the multi-pollutant concentration prediction method based on the spatio-temporal attention and the gated reset recurrent network.
[0018] Compared with the prior art, the application has the beneficial effects that: the application, through a joint modeling framework of residual convolution coding, double spatio-temporal attention, and attention-enhanced GRU (containing Attn-Reset), significantly reduces prediction errors and improves the determination coefficient R2 in single-step and 3 / 6 / 12 / 24 h multi-step look-ahead scenarios, and has high precision and stability; relying on 0.1-5 km multi-scale concentric buffer zones for the fusion modeling of traffic and land use characteristics, taking into account local and regional driving factors, and enhancing multi-scale representation capability; introducing a dynamic reset gate to effectively suppress long-term error accumulation and improve the robustness and stability of multi-step prediction; adopting modular design and clear training and inference process, facilitating integration and deployment with existing urban environment monitoring platforms, and having good engineering feasibility. BRIEF DESCRIPTION OF DRAWINGS
[0019] Figure 1 FIG. 1 is a schematic diagram of the overall process of the application based on the spatio-temporal attention and the gated reset recurrent network (STAGR-Net); Figure 2 FIG. 2 is a system architecture diagram of the application based on the spatio-temporal attention and the gated reset recurrent network (STAGR-Net); Figure 3 FIG. 3 is a schematic diagram of the Attention-GRU structure. DETAILED DESCRIPTION
[0020] The application will be further described below in combination with the drawings and specific embodiments.
[0021] Embodiment 1: The embodiment provides a multi-pollutant concentration prediction method based on the spatio-temporal attention and the gated reset recurrent network, comprising the following steps: S1. Data Acquisition and Multi-Scale Alignment At an hourly time resolution, pollutant concentration sequences (CO, NO, NO2, O3, BC, PM) were obtained. 2.5 PM 10 Meteorological elements (air pressure, dew point, temperature, humidity, precipitation, visibility, wind speed, wind direction), hourly traffic indicators based on 15-minute traffic flow, and static land use attributes (residential, commercial, industrial, green space, transportation facilities, etc.). The spatial representation of traffic flow characteristics involves constructing multi-scale concentric buffer zones of 0.1–5 km with a step size of 0.1 km around the monitoring stations, extracting and aggregating multi-scale traffic / land use features, and constructing time-social indicator variables such as weekdays / weekends, lockdowns or holidays, and heating seasons.
[0022] The above multi-source features are aligned and registered at all time steps and spatial scales to form a spatiotemporal feature matrix of dimension [W×F×S], where W is the length of the time window, F is the feature dimension, and S is the number of spatial scales.
[0023] S2. Perform multi-layer residual convolution feature extraction on the aligned multi-source features to obtain high-level semantic representations. The preferred approach is to use a combination of 3×3 convolutions with skip connections, normalization, and activation to enhance the representation of boundaries and local structures and suppress gradient vanishing.
[0024] S3, a hybrid encoding of spatiotemporal attention and residual convolution, will The input is fed into the spatiotemporal attention module, which adaptively weights key regions and key time periods and extracts robust representations to obtain a [W×F] hybrid feature map, including: right Perform max pooling and average pooling along the time dimension respectively to obtain The spatial weight map is obtained by concatenating the two along the channel dimension and then performing a 3×3 convolution followed by a Sigmoid gate. ,according to Features are weighted to highlight traffic hotspots and weather-sensitive areas, among which This indicates element-wise multiplication.
[0025] Will The mapping yields query Q, key K, and value V to emphasize abrupt, lag, and periodic patterns, calculates attention weights, and obtains temporal attention features. ,
[0026] ; In the formula: d is the scaling factor.
[0027] fusion and get spatiotemporal weighted features , input residual blocks composed of multi-layer 3x3 convolution, normalization and activation, stabilize gradients and enhance boundary and local structure expression through skip connection, output mixed feature map of [WxF] for sequence modeling.
[0028] S4, input spatiotemporal weighted features into attention-enhanced gated recurrent unit, introduce dynamic reset mechanism (Attn-Reset) sensitive to attention weight in GRU reset gate to adaptively adjust memory / forgetting between traffic peak mutation and stable stage, output final hidden state sequence ; attention weight is injected into GRU reset gate as gate bias or scaling item, making reset gate adaptively sensitive to key time slices to achieve state switching between rapid transition of morning and evening traffic peaks and background stable stage, thereby suppressing error propagation and enhancing stability of multi-step prediction.
[0029] As shown in Figure 3 , attention-enhanced gated recurrent unit introduces modulation of attention weight (dynamic reset mechanism, Attn-Reset) at the reset gate and candidate state gate of standard GRU to adaptively switch memory / forgetting strength between rapid transition and slow evolution, specifically:
[0030]
[0031]
[0032]
[0033] where is Sigmoid, is hyperbolic tangent, is scaling / bias mapping of attention weight; , , are trainable parameters.
[0034] S5, for time step prediction, input hidden state sequence into fully connected layer to get one-step prediction of next hour; apply Dropout at hidden or output layer during training period to alleviate overfitting.
[0035] For multi-step forward-looking prediction, direct or time-shifted methods are used. The direct method uses stacked fully connected layers to output 3h, 6h, 12h, and 24h prediction values in parallel in the same forward process. The time-shifted method generates multi-look-ahead results sequentially through a recursive or time-shift mechanism. Dropout is applied to the hidden or output layer to alleviate overfitting.
[0036] S6, training target and optimization strategy: the training stage uses Smooth L1 loss, uses Adam optimizer and combines cosine annealing learning rate scheduling for training to improve convergence stability and generalization ability.
[0037] Embodiment 2: The embodiment provides a multi-pollutant concentration prediction system based on a space-time attention and a gated reset recurrent network, comprising: A data acquisition and alignment module is configured to acquire and align pollutants, weather, traffic flow, and land use multi-source features. A residual convolutional encoding module is configured to perform multi-layer residual convolutional feature extraction on the aligned multi-source features. A space-time attention module is configured to adaptively weight and fuse key areas and key time periods of the extracted features to obtain space-time weighted features, and to obtain a mixed feature mapping by convolving the space-time weighted features. The space-time attention module includes a spatial attention submodule and a temporal attention submodule. The spatial attention submodule applies max-pooling and average-pooling along the time dimension, concatenates the features in the channel dimension, and then generates a spatial weight map through 3x3 convolution and Sigmoid activation. The spatial weight map is then used to weight the features to generate spatial attention features. The temporal attention submodule maps the features to query Q, key K, and value V, calculates attention weights, and outputs temporal attention features.
[0038] An attention-enhanced gated recurrent module is configured to input the mixed feature mapping by time step to generate a hidden state sequence. A prediction and regularization module is configured to output a time-step or multi-step forward-looking prediction result based on the hidden state sequence. A training and optimization module is configured to use Smooth L1 as a loss function, use Adam optimizer, and combine cosine annealing learning rate scheduling for end-to-end training. A result output module is configured to output prediction results for each pollutant.
[0039] The above description is only a preferred embodiment of the present application and does not limit the present application. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A multi-pollutant concentration prediction system based on spatiotemporal attention and gating reset loop network, characterized in that: include: The data acquisition and alignment module is used to acquire and align multi-source characteristics of pollutants, meteorology, traffic flow, and land use. The residual convolutional coding module is used to extract aligned multi-source features through multi-layer residual convolutional features; The spatial-temporal attention module is used to adaptively weight and fuse key regions and key time periods for feature extraction to obtain spatiotemporally weighted features, and then convolves the spatiotemporally weighted features to obtain a hybrid feature map. The attention-enhanced gating loop module is used to generate a sequence of hidden states by taking the hybrid feature map as input over time steps; The prediction and regularization module is used to output time-step or multi-step look-ahead prediction results based on the hidden state sequence; The training and optimization module is used for end-to-end training with Smooth L1 loss function, Adam optimizer and cosine annealing learning rate scheduling. The results output module is used to output the prediction results for each pollutant.
2. The multi-pollutant concentration prediction system based on spatiotemporal attention and gating reset loop network according to claim 1, characterized in that: The space-time attention module includes a spatial attention submodule and a temporal attention submodule; The spatial attention submodule extracts features Max pooling and average pooling are applied along the time dimension, and after concatenation in the channel dimension, a spatial weight map is generated by 3×3 convolution and sigmoid activation. Then, the features are weighted based on the spatial weight map to generate spatial attention features. The time attention submodule maps features to query Q, key K, and value V, calculates attention weights, and outputs time attention features.
3. A multi-pollutant concentration prediction method based on spatiotemporal attention and gating reset recurrent networks, characterized in that: Includes the following steps: S1. Obtain the pollutant concentration sequence, meteorological elements, traffic flow indicators and static land use attribute characteristics of the target area at hourly time resolution, and align the features with spatial location at all time steps. S2. Perform multi-layer feature extraction on the obtained alignment features to obtain a high-level semantic representation. ; S3, Based on high-level semantic representation The spatiotemporal weighted features are obtained by fusing key spatial regions and key time segments. Then, the input consists of residual blocks composed of multiple 3×3 convolutions, normalization, and activation, and the output is a hybrid feature map; S4. Input the hybrid feature map into the attention-enhanced gated recurrent module at time steps to generate a hidden state sequence; S5. Input the hidden state sequence into the prediction and regularization module and output the time-step or multi-step look-ahead prediction results. S6. Using Smooth L1 loss as the training objective, the Adam optimizer is employed in conjunction with cosine annealing learning rate scheduling for training optimization.
4. The multi-pollutant concentration prediction method based on spatiotemporal attention and gating reset recurrent network according to claim 3, characterized in that: The pollutants mentioned in step S1 include CO, NO, NO2, O3, BC, and PM. 2.5 PM 10 The meteorological elements include air pressure, dew point, temperature, humidity, precipitation, visibility, wind speed, and wind direction; the static land use attributes include residential, commercial, industrial, green space, and transportation facilities.
5. The multi-pollutant concentration prediction method based on spatiotemporal attention and gating reset recurrent networks according to claim 3, characterized in that: The spatial representation of traffic flow characteristics in S1 includes: constructing multi-scale concentric buffer zones of 0.1-5 km with a step size of 0.1 km around each monitoring station, extracting and aggregating the overall level and spatial difference characteristics of traffic flow, and aligning them one by one with the hourly time series input according to the timestamp.
6. The multi-pollutant concentration prediction method based on spatiotemporal attention and gating reset recurrent network according to claim 3, characterized in that: Step S3 includes: (1) Features Max pooling and average pooling are applied along the time dimension, respectively. ； ； Will and After concatenation of the channel dimensions, a spatial weight map is generated by 3×3 convolution and sigmoid activation. Then, spatial attention features are generated using weighted methods. , ， In the formula: This represents element-wise multiplication; (2) Features The mapping is defined as query Q, key K, and value V. Attention weights are calculated, and the temporal attention features are output using these weights. , ； ； In the formula: d is the scaling factor; (3) Integration and obtain spatiotemporal weighted features and spatiotemporal weighted features The input consists of residual blocks composed of multiple 3×3 convolutions, normalization, and activation, and the output is a hybrid feature map.
7. The multi-pollutant concentration prediction method based on spatiotemporal attention and gating reset recurrent network according to claim 3, characterized in that: In step S4, the dynamic reset gate in the attention enhancement gating loop module injects the attention weights as gating biases or scaling terms into the GRU reset gate.
8. The multi-pollutant concentration prediction system based on spatiotemporal attention and gating reset loop network according to claim 3, characterized in that: In step S5, for time step prediction, the prediction and regularization module inputs the hidden state sequence into the fully connected layer to obtain the next hour's one-step prediction, and applies Dropout to the hidden or output layer to alleviate overfitting.
9. The multi-pollutant concentration prediction system based on spatiotemporal attention and gating reset loop network according to claim 3, characterized in that: In step S5, for multi-step prediction, the prediction and regularization modules use stacked fully connected layers to output 3h, 6h, 12h, and 24h prediction values in parallel during the same forward process; or generate multiple forward results sequentially through recursion or time shifting mechanisms.
10. A readable storage medium storing program instructions, wherein a computer program is stored thereon, characterized in that, When executed by the processor, the program implements the multi-pollutant concentration prediction method based on spatiotemporal attention and gating reset loop network as described in any one of claims 3-9.