A multi-region environment monitoring data coupling time interval deduction method and system

Abstract

The application discloses a kind of multi-region environmental monitoring data coupling time sequence interval deduction method and system, it is related to time sequence data processing technical field.The method obtains the historical target interval sequence of regional environmental monitoring data unit, environmental monitoring data and data processing specification text, and obtains real target interval sequence and daily environmental characteristics after pre-processing;Based on dynamic time warping distance, construct regional correlation diagram, generate double-layer environmental factors using time-varying environmental stress index to adjust daily environmental characteristics;Fusion specification text semantic vector and time-varying environmental stress index into conditional vector, generate extreme state target interval sequence through conditional generative adversarial network;Real training sample and extreme state training sample and sample label are constructed, and enhanced training sample set is obtained;Multi-scale decomposition is carried out to training sample input window, and multi-scale graph space-time network is trained in combination with regional correlation diagram and double-layer environmental factors, and interval deduction result containing deduction upper bound and deduction lower bound is output.

Description

Technical Field

[0001] This invention relates to the field of multi-source time-series data processing technology, specifically to a time-series interval extrapolation method and system for coupling multi-regional environmental monitoring data. Background Technology

[0002] In multi-regional environmental monitoring, carbon emission monitoring, and pollutant emission monitoring scenarios, different regions typically continuously collect time-series data such as carbon emissions, pollutant concentrations, meteorological and environmental conditions, energy consumption, or equipment operating load. Due to factors such as the accuracy of monitoring equipment, data sampling frequency, environmental disturbances, data gaps, and differences in statistical methods, monitoring results at the same point in time or within the same statistical period often have a certain degree of uncertainty, and usually need to be expressed in the form of an interval composed of upper and lower boundary values.

[0003] Existing time-series extrapolation methods typically process monitoring data from different regions separately or establish correlations based on static spatial distances between regions. This makes it difficult to accurately reflect the response lags, inconsistent change rhythms, and phased linkages between target interval sequences in different regions. Furthermore, the degree of influence of environmental factors such as temperature, humidity, wind speed, precipitation, and pollutant concentrations on monitoring results in different regions changes over time. Using fixed weights to process environmental characteristics is insufficient to adapt to dynamic changes in environmental pressure.

[0004] Furthermore, extreme conditions such as extreme weather, abnormal pollutant concentrations, or high equipment load occur infrequently, resulting in a limited number of actual extreme condition samples collected, which can easily lead to an unbalanced distribution of training samples. Existing methods also struggle to incorporate the data processing specifications, monitoring limits, equipment calibration rules, anomaly detection rules, and quality control requirements contained in the data processing guidelines into the generation of extreme condition samples and interval extrapolation, leading to insufficient stability and robustness of the model under extreme conditions. Therefore, a time-series interval extrapolation method that can comprehensively consider multi-regional temporal correlations, dynamic environmental disturbances, data processing specifications, and the problem of insufficient extreme condition samples is needed. Summary of the Invention

[0005] To address the problems in existing technologies, such as the difficulty in accurately representing temporal misalignments between environmental monitoring data from different regions, the inability to dynamically adjust environmental feature weights according to environmental pressure states, the difficulty in incorporating data processing specifications into model processing, and the poor stability and robustness of interval extrapolation due to insufficient extreme state samples, this invention provides a time-series interval extrapolation method and system that couples multi-regional environmental monitoring data. This method constructs a regional correlation graph and a two-layer environmental factor, utilizes a conditional generative adversarial network to generate extreme state target interval sequences, and constructs an enhanced training sample set based on real training samples, extreme state training samples, and their labels. It then performs joint training using multi-scale decomposition and a multi-scale graph spatiotemporal network to output interval extrapolation results containing the upper and lower boundaries of the extrapolation. To achieve the above objectives, this invention proposes a time-series interval extrapolation method that couples multi-regional environmental monitoring data, comprising the following steps: step S 1. Obtain historical target interval sequences, environmental monitoring data, and data processing specification texts from multiple regional environmental monitoring data units and perform preprocessing to obtain the actual target interval sequences and daily environmental characteristics, respectively; wherein, the historical target interval sequences are interval-based operational status data collected or generated by regional environmental monitoring data units; step S 2: Calculate the dynamic time-warped distance based on the real target interval sequence to construct a regional correlation map; calculate the time-varying environmental pressure index based on the daily environmental characteristics, and use the time-varying environmental pressure index to adjust the weight of the daily environmental characteristics to generate a two-layer environmental factor; step S 3: Semantically encode the data processing standard text to obtain a standard text semantic vector. Fuse the standard text semantic vector with the time-varying environmental pressure index to form a condition vector. Input the condition vector into a conditional generative adversarial network for adversarial training to generate an extreme state target interval sequence. step S 4. Divide the real target interval sequence according to the input window length and the inference step size to construct real training samples and their sample labels; divide the extreme state target interval sequence according to the same input window length and inference step size to construct extreme state training samples and their sample labels; mix the real training samples with the extreme state training samples to obtain an enhanced training sample set. step S 5: Decompose the input window of each training sample in the enhanced training sample set to obtain multi-scale interval features, and input the multi-scale interval features, the region association map and the two-layer environmental factors into the initial multi-scale graph spatiotemporal network. Use the sample labels as supervision information to perform joint optimization training to obtain the trained multi-scale graph spatiotemporal network. step S 6: Input the current multi-scale interval features, the current regional association map, and the current two-layer environmental factors corresponding to the time period to be inferred into the trained multi-scale graph spatiotemporal network for forward inference to obtain the interval inference result including the upper and lower boundaries of the inference.

[0006] Preferably, the steps S In step 2, the dynamic time-warped distance is calculated based on the real target interval sequence to construct a regional association graph, specifically including: Calculate the first The region and the first The dynamic time-warped distance between the real target interval sequences of each region is used to obtain a distance matrix; wherein, the real target interval sequence is a two-dimensional interval vector sequence composed of upper and lower boundary values, and the dynamic time-warped distance is calculated based on the two-dimensional interval vector sequence; The calculation of the dynamic time warping distance includes: determining the local distance between binary interval vectors based on the difference between the upper and lower boundaries of two regions at different time points, and determining the dynamic time warping distance based on the local distance according to the minimum cumulative distance of the dynamic time warping path; The distance matrix is ​​converted into edge weights using a distance scale parameter; Using multiple regional environmental monitoring data units as nodes and the edge weights as elements of an adjacency matrix, an undirected weighted graph is constructed as the regional association graph. The formula for calculating the edge weights is as follows:

[0007] in, For the first The region and the first Edge weights between regions For the first The region and the first Dynamic time-normalized distance between regions For distance scale parameters, and .

[0008] Preferably, the steps S In step 2, a time-varying environmental pressure index is calculated based on the daily environmental characteristics, and the weights of the daily environmental characteristics are adjusted using the time-varying environmental pressure index to generate a two-layer environmental factor, specifically including: The time-varying environmental pressure index is obtained by weighted summing of the extreme weather event frequency index, the proportion of days exceeding air quality standards, and the comprehensive pollutant concentration index. The calculation formula is as follows:

[0009] in, For the first The time-varying environmental stress index for each statistical period. For the first The frequency index of extreme weather events over a statistical period For the first The percentage of days with air quality exceeding standards for each statistical period. For the first The comprehensive index of pollutant concentrations for each statistical period. , and For the corresponding component weights, and , and All are not less than 0. ; The time-varying environmental pressure index is input into the environmental gating unit to generate gating weights, and the gating weights are then multiplied element-wise with the daily environmental characteristics to obtain enhanced environmental characteristics. The time-varying environmental pressure index and the enhanced environmental characteristics together constitute the two-layer environmental factor, and the calculation formula for the enhanced environmental characteristics is as follows:

[0010]

[0011] in, For the first The gating weights corresponding to each daily time point For activation function, and For learnable parameters, Indicates the first The statistical period to which each daily time point belongs. For the first The time-varying environmental stress index corresponding to the statistical period of each daily time point. For the first Daily environmental characteristics at a specific time of day. For the first Enhanced environmental characteristics at individual daily time points This is an element-wise multiplication operation; the gating weight is a scalar or a vector with the same dimension as the daily environmental feature.

[0012] Preferably, the conditional generative adversarial network includes a generator and a discriminator; The steps S 3 also includes: Obtain historical data processing specification text, perform word segmentation, stop word removal, and vectorization on the historical data processing specification text, and use word vector models, pre-trained language models, or... ransfor er The encoding model extracts semantic features to obtain the semantic vector of the standard text; The conditional vector is obtained by concatenating the standard text semantic vector with the time-varying environmental pressure index. The conditional vector and random noise are input into the generator, which outputs the corresponding target interval sequence. The generated target interval sequence and the condition vector are input together into the discriminator, and the real target interval sequence and the condition vector are input together into the discriminator, which then performs an authenticity evaluation. After the generator and the discriminator are trained adversarially until the changes in both the generator loss and the discriminator loss are less than a preset loss threshold, or after the number of training rounds reaches a preset number of rounds, the generator output target interval sequence is extracted. The generated target interval sequence is subjected to boundary consistency correction to ensure that the upper boundary of generation is not lower than the lower boundary of generation. Then, the generated target interval sequence with time-varying environmental pressure index greater than a preset pressure threshold or target interval fluctuation amplitude greater than a preset fluctuation threshold is selected from the generated target interval sequence after boundary consistency correction to obtain the extreme state target interval sequence.

[0013] Preferably, the steps S 4. Specifically includes: According to the preset input window length and inference step size, the real target interval sequence is divided into multiple real training samples. Each real training sample includes an input window consisting of real target intervals at multiple consecutive time points, and a sample label located after the input window as supervision information. The extreme state target interval sequence is divided into multiple extreme state training samples according to the same input window length and inference step size. Each extreme state training sample includes an input window consisting of extreme state target intervals at multiple consecutive time points, and a sample label located after the input window as supervision information. The real training samples and the extreme state training samples are mixed according to a preset mixing ratio to obtain the enhanced training sample set.

[0014] Preferably, the steps S In step 5, the input window of each training sample in the enhanced training sample set is decomposed to obtain multi-scale interval features, specifically including: Bivariate empirical mode decomposition is used to jointly decompose the upper and lower boundary sequences in the input window, separating multiple intrinsic mode function components and a residual term respectively; All the intrinsic mode function components and all the residual terms are used together as the multi-scale interval features; The calculation formula for the joint decomposition is:

[0015]

[0016] in, and The input window at time step The upper and lower boundaries, This represents the total number of intrinsic mode function components. and The first The components corresponding to the upper and lower boundaries among the eigenmode function components, and These are the residual terms corresponding to the upper and lower boundaries, respectively.

[0017] Preferably, the multi-scale graph spatiotemporal network includes a graph convolutional long short-term memory network layer and a spatiotemporal collaborative attention layer connected in sequence; The steps S 5 specifically includes: Using the graph convolutional long short-term memory network layer based on the region association graph, spatial neighbor feature aggregation and long-term temporal dependency extraction are performed on the multi-scale interval features and the two-layer environmental factors to obtain basic spatiotemporal features. The multi-head self-attention mechanism in the spatiotemporal collaborative attention layer is used to adaptively allocate the weights of environmental monitoring data units in each region at different time points, and the basic spatiotemporal features are weighted and mapped into higher-order spatiotemporal features. Based on the higher-order spatiotemporal features, the upper and lower bounds of the inference are output, and the inference error is calculated according to the upper bound, the lower bound and the corresponding sample labels. The network weights of the multi-scale graph spatiotemporal network are then updated by backpropagation.

[0018] Preferably, the joint optimization training uses a multi-objective loss function to calculate the inference error, and the multi-objective loss function includes numerical error loss, directional consistency loss and interval consistency loss; The numerical error loss is used to constrain the upper and lower boundaries of the inference to be close to the upper and lower boundaries of the labels in the sample labels, respectively. The directional consistency loss is used to constrain the direction of change of the center value of the extrapolation interval to be consistent with the direction of change of the center value of the label interval corresponding to the sample label; The interval consistency loss is used to constrain the upper boundary of the deduction to be no lower than the lower boundary of the deduction; the interval consistency loss includes a boundary penalty term generated when the lower boundary of the deduction is higher than the upper boundary of the deduction. Wherein, the center value of the deduction interval is the average of the upper and lower boundaries of the deduction, and the center value of the label interval is the average of the upper and lower boundaries of the label.

[0019] Preferably, the steps S 6 specifically includes: The historical target interval window before the period to be simulated is arranged according to the steps. S 5. Process the data using the same decomposition method to obtain the current multi-scale interval features; The current regional association map is obtained by updating or reusing the actual target interval sequence before the time period to be simulated. Obtain environmental monitoring data corresponding to the time period to be simulated, obtain daily environmental characteristics corresponding to the time period to be simulated based on the environmental monitoring data, calculate the time-varying environmental pressure index corresponding to the time period to be simulated based on the daily environmental characteristics, adjust the weight of the daily environmental characteristics using the time-varying environmental pressure index, and generate the current two-layer environmental factor. The current multi-scale interval features, the current region association map, and the current two-layer environmental factors are jointly input into the trained multi-scale graph spatiotemporal network. The multi-scale graph spatiotemporal network, after training, outputs the upper and lower boundaries of the time period to be inferred, and uses the numerical range formed by the upper and lower boundaries as the interval inference result.

[0020] A time-series inter-regional extrapolation system coupling multi-regional environmental monitoring data includes: The data processing module is configured to acquire historical target interval sequences, environmental monitoring data, and data processing specification texts from multiple regional environmental monitoring data units and perform preprocessing to obtain the actual target interval sequences and daily environmental characteristics, respectively. The graph and factor construction module is configured to calculate the dynamic time-warped distance based on the real target interval sequence to construct a regional association graph, calculate the time-varying environmental pressure index based on the daily environmental characteristics, and use the time-varying environmental pressure index to adjust the weight of the daily environmental characteristics to generate a two-layer environmental factor. The adversarial generation module is configured to semantically encode the data processing standard text to obtain a standard text semantic vector, fuse the standard text semantic vector with the time-varying environmental pressure index to form a conditional vector, and input the conditional vector into a conditional generative adversarial network for adversarial training to generate an extreme state target interval sequence. The sample construction module is configured to divide the real target interval sequence and the extreme state target interval sequence according to the input window length and the inference step size, respectively, to construct real training samples, extreme state training samples and corresponding sample labels, and to mix the real training samples and the extreme state training samples to obtain an enhanced training sample set. The network training module is configured to decompose the input window of each training sample in the enhanced training sample set to obtain multi-scale interval features, and input the multi-scale interval features, the region association graph and the two-layer environmental factors into the initial multi-scale graph spatiotemporal network, and perform joint optimization training with the sample labels as supervision information to obtain the trained multi-scale graph spatiotemporal network. The inference module is configured to input the current multi-scale interval features, the current regional association map, and the current two-layer environmental factors corresponding to the time period to be inferred into the trained multi-scale graph spatiotemporal network for forward inference, and output the interval inference result including the upper and lower boundaries of the inference.

[0021] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a regional correlation graph based on the dynamic time-warped distance of real target interval sequences, which can characterize the dynamic correlation between different regions with temporal misalignment. It utilizes a time-varying environmental pressure index to adjust the weights of daily environmental features and generate a two-layer environmental factor, improving the model's adaptability to dynamic environmental disturbances. By fusing the normative text semantic vector with the time-varying environmental pressure index to form a conditional vector, and generating extreme-state target interval sequences through a conditional generative adversarial network, it constructs an enhanced training sample set containing real training samples and extreme-state training samples, alleviating the problems of insufficient extreme-state samples and unbalanced training sample distribution. Through multi-scale decomposition and multi-scale graph spatiotemporal networks jointly processing regional correlation, time dependence, and environmental disturbance information, and jointly constraining the upper and lower bounds of the extrapolation, it reduces the risk of interval boundary inversion and abnormal fluctuations, improving the stability and robustness of time-series interval extrapolation of multi-regional environmental monitoring data. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the time-series interval extrapolation method for coupling multi-regional environmental monitoring data according to the present invention.

[0023] Figure 2 This is a schematic diagram of the processing flow of the multi-scale graph spatiotemporal network of the present invention.

[0024] Figure 3 This is a schematic diagram of the processing flow of the conditional generative adversarial network of the present invention.

[0025] Figure 4 This is an embodiment of the present invention. PM2.5 Comparison chart of concentration range extrapolation results.

[0026] Figure 5 This is an embodiment of the present invention. PM 2.5 Absolute error diagram of concentration range boundaries.

[0027] Figure 6 This is a comparison chart showing the results of carbon emission range extrapolation in embodiments of the present invention.

[0028] Figure 7 This is a graph showing the absolute error of the carbon emission range boundaries in an embodiment of the present invention.

[0029] Figure 8 This is a schematic diagram of the system architecture of the time-series interval extrapolation system of the present invention. Detailed Implementation

[0030] like Figures 1 to 4 As shown, in this embodiment, the historical target interval sequence refers to the interval-based operational status data with upper and lower boundary values, formed by the collection, statistics, or calculation of multiple regional environmental monitoring data units at historical time points; the real target interval sequence refers to the interval sequence obtained after time alignment, missing value imputation, outlier processing, and normalization of the historical target interval sequence; the daily environmental characteristics refer to meteorological data, pollutant concentration data, or equipment operation status data matched with the corresponding time point; the two-layer environmental factor refers to the model input factor composed of the time-varying environmental pressure index and the enhanced environmental characteristics adjusted by gating weights; the extreme state target interval sequence refers to the target interval sequence generated or screened under high environmental pressure conditions or large fluctuations in the target interval; the enhanced training sample set refers to the training sample set formed by mixing real training samples and extreme state training samples, and each training sample includes an input window and its corresponding sample label.

[0031] like Figure 1 As shown, this embodiment provides a time-series interval extrapolation method for coupling multi-regional environmental monitoring data. This method relies on underlying hardware devices for execution and includes the following steps: S 1 to S 6.

[0032] step S 1: Obtain historical target interval sequences, environmental monitoring data, and data processing specification texts from multiple regional environmental monitoring data units and perform preprocessing to obtain corresponding real target interval sequences and daily environmental characteristics. The real target interval sequences include upper boundary sequences and lower boundary sequences.

[0033] The plurality of regional environmental monitoring data units include at least one of the following: environmental monitoring data units deployed in different regions, carbon emission monitoring data units, energy operation data acquisition units, or pollutant monitoring data units. The historical target interval sequence is intervalized operational status data collected, statistically analyzed, or calculated by the plurality of regional environmental monitoring data units at multiple historical time points, and the intervalized operational status data includes upper boundary values ​​and lower boundary values.

[0034] The historical target interval sequence includes at least one of the following: regional carbon emission interval, carbon emission intensity interval, energy consumption intensity interval, pollutant concentration interval, pollutant emission intensity interval, environmental quality index interval, or equipment operating load interval; the upper boundary value and the lower boundary value are used to characterize the upper limit estimate and lower limit estimate of the corresponding operating status data at the same point in time or within the same statistical period, respectively.

[0035] Specifically, target interval data and auxiliary time-series data formed by multiple regional environmental monitoring data units at historical time points are collected. The target interval data includes upper and lower boundary values, and the auxiliary time-series data includes at least one of the following: initial monitoring value, final monitoring value, monitoring value change range, cumulative emissions, cumulative energy consumption, equipment operating load, or data acquisition frequency. Simultaneously, meteorological data and pollution data corresponding to each regional environmental monitoring data unit are collected. The meteorological data includes at least one of temperature, humidity, wind speed, and precipitation, and the pollution data includes… PM 2.5 PM 10. NO 2. SO 2 or O At least one of the three concentrations.

[0036] Time alignment of the above data, n - ax Normalization, missing value imputation, and outlier handling ultimately form the true target interval sequence and daily environmental characteristics. Missing value imputation can be achieved by interpolating between adjacent time points, imputing with the mean of the same period, or imputing with preset missing markers; outlier handling can be achieved by threshold removal, upper and lower boundary truncation, or correction based on historical statistical ranges.

[0037] step S 2: Calculate the dynamic time-normalized distance based on the real target interval sequence to construct a regional correlation map; calculate the time-varying environmental pressure index based on the daily environmental characteristics, and use the time-varying environmental pressure index to adjust the weight of the daily environmental characteristics to generate a two-layer environmental factor.

[0038] Constructing the regional association graph includes: Since traditional static distance cannot measure the temporal misalignment association of the target interval sequence, this embodiment calculates the first... The region and the first The distance matrix is ​​obtained by calculating the dynamic time-warped distance between the real target interval sequences of each region. The real target interval sequence is a two-dimensional interval vector sequence composed of upper and lower boundary values, and the dynamic time-warped distance is calculated based on this two-dimensional interval vector sequence.

[0039] In calculating the first The region and the first When calculating the dynamic time-normalized distance between regions, the first region will be... The region in the first The actual target interval at each time point is represented as a binary interval vector:

[0040] The first The region in the first The actual target interval at each time point is represented as a binary interval vector:

[0041] in, and They represent the first The region in the first The lower and upper boundary values ​​at each time point and They represent the first The region in the first The lower and upper boundary values ​​at each time point. The local distance between two binary interval vectors is:

[0042] Construct a local distance matrix based on the local distances, and obtain the first local distance matrix according to the principle of minimizing the cumulative distance of the dynamically time-warped path. The region and the first Dynamic time warping distance between regions .

[0043] The distance matrix is ​​converted into edge weights using a distance scale parameter. Multiple regional environmental monitoring data units are used as nodes, and the edge weights are used as elements of the adjacency matrix to construct an undirected weighted graph as a regional association graph. The formula for calculating the edge weights is as follows:

[0044] in, For the first The region and the first Edge weights between regions For the first The region and the first Dynamic time-normalized distance between regions The distance scale parameters are determined through grid search, and .

[0045] The two-layer environmental factor processing method includes: weighting and summing the extreme weather event frequency index, the proportion of days exceeding air quality standards, and the comprehensive pollutant concentration index to obtain the time-varying environmental pressure index.

[0046] in, For the first Time-varying environmental stress index for each statistical period For the first The frequency index of extreme weather events over a statistical period For the first The percentage of days with air quality exceeding standards for each statistical period. For the first The comprehensive index of pollutant concentrations for each statistical period. , and For the corresponding component weights, and , and All are not less than 0. .

[0047] Extreme weather event frequency index The proportion of days with air quality exceeding standards can be determined by the ratio of the number of days with extreme high temperatures, heavy precipitation, and strong winds to the total number of days in the statistical period. The comprehensive pollutant concentration index can be determined based on the ratio of the number of days when air quality indicators exceed preset limits to the total number of days in the statistical period; It can be determined by weighted summation of the normalized values ​​of multiple pollutant concentrations relative to the corresponding limits.

[0048] Subsequently, the time-varying environmental pressure index is input into the environmental gating unit to generate gating weights, which dynamically adjust the channel importance of daily environmental characteristics:

[0049]

[0050] in, For the first The gating weights corresponding to each daily time point For activation function, and For learnable parameters, Indicates the first The statistical period to which each daily time point belongs. For the first The time-varying environmental stress index corresponding to the statistical period of each daily time point. For the first Daily environmental characteristics at a specific time of day. For the first Enhanced environmental characteristics at individual daily time points This is an element-wise multiplication operation. The gating weights are scalars or vectors with the same dimension as the daily environmental features.

[0051] The time-varying environmental pressure index and the enhanced environmental feature together constitute a two-layer environmental factor. The time-varying environmental pressure index is used to reflect the current environmental pressure state from a statistical period perspective, while the enhanced environmental feature is used to reflect the environmental change characteristics after gating adjustment from a daily time point perspective.

[0052] step S 3: Semantically encode the data processing standard text to obtain a standard text semantic vector. Fuse the standard text semantic vector with the time-varying environmental pressure index to form a condition vector. Input the condition vector into a conditional generative adversarial network for adversarial training to generate an extreme state target interval sequence.

[0053] Specifically, such as Figure 3 As shown, historical data processing specification texts related to the operational status of regional environmental monitoring data units are obtained. These historical data processing specification texts include at least one of the following: data acquisition specification texts, equipment calibration specification texts, pollutant monitoring limit texts, carbon emission accounting technical specification texts, abnormal data judgment rule texts, or data quality control rule texts. A pre-trained Chinese language model is used. BER - ww The data is encoded as a fixed-dimensional semantic feature, such as a 768-dimensional canonical text semantic vector. This canonical text semantic vector is then concatenated with a 1-dimensional time-varying environmental pressure index to obtain a 769-dimensional conditional vector. This canonical text semantic vector is used to characterize the data collection criteria, limit conditions, anomaly detection conditions, and quality control requirements, and serves as the conditional input information when generating the target interval sequence.

[0054] Conditional Generative Adversarial Networks (GANs) consist of a generator and a discriminator. The generator employs a three-layer architecture. LSTM The structure receives a concatenated input of random noise and a conditional vector, and outputs a target interval sequence; the discriminator employs a 3-layer 1... D - CNNStructure. The generated target interval sequence and the condition vector are input together into the discriminator, and the real target interval sequence and the condition vector are also input together into the discriminator, which then performs an authenticity evaluation.

[0055] The generator is trained to make the discriminator classify the generated target interval sequence as a real sequence, while the discriminator is trained to distinguish between the real target interval sequence and the generated target interval sequence. The generator and discriminator are trained adversarially until the changes in the loss of the discriminator and the loss of the generator are both less than a preset loss threshold, or the preset number of training rounds is reached.

[0056] The generator outputs a target interval sequence including a lower boundary sequence and an upper boundary sequence corresponding to the target time step. After obtaining the target interval sequence, boundary consistency correction is performed to ensure that the upper boundary is not lower than the lower boundary. Furthermore, based on whether the time-varying environmental pressure index is greater than a preset pressure threshold and whether the target interval fluctuation amplitude is greater than a preset fluctuation threshold, the target interval sequence after boundary consistency correction is filtered to obtain the extreme state target interval sequence, thereby alleviating the problem of insufficient real extreme state samples.

[0057] step S 4. Divide the real target interval sequence according to the input window length and inference step size to construct real training samples and their labels; divide the extreme state target interval sequence according to the same input window length and inference step size to construct extreme state training samples and their labels; mix the real training samples and the extreme state training samples to obtain the enhanced training sample set.

[0058] Specifically, according to the preset input window length, multiple consecutive time points of real target intervals are sequentially extracted from the real target interval sequence to form the input window of the real training sample; the real target interval located after the input window and corresponding to the preset inference step size is used as the sample label of the real training sample.

[0059] Using the same input window length and inference step size, extreme state target intervals at multiple consecutive time points are extracted from the extreme state target interval sequence to form the input window of the extreme state training sample; the extreme state target interval located after the input window and corresponding to the preset inference step size is used as the sample label of the extreme state training sample.

[0060] Real training samples and extreme state training samples are mixed according to a preset mixing ratio to obtain an enhanced training sample set. The preset mixing ratio can be determined based on the proportion of extreme state training samples to the number of samples in a single training batch, for example, making extreme state training samples account for 10%-40% of the number of samples in a single training batch. By adding extreme state training samples to real training samples, the distribution of the number of normal and extreme states in the training samples can be improved without changing the correspondence between the input window and the sample label within the real training samples.

[0061] like Figure 2 As shown, the steps S 5: Decompose the input window of each training sample in the enhanced training sample set to obtain multi-scale interval features, and input the multi-scale interval features, the region association map and the two-layer environmental factors into the initial multi-scale graph spatiotemporal network. Use the sample labels as supervision information to perform joint optimization training to obtain the trained multi-scale graph spatiotemporal network.

[0062] Bivariate Empirical Mode Decomposition BEMD Perform joint decomposition on the upper and lower boundary sequences in the input window to preserve the phase information between them:

[0063]

[0064] All intrinsic mode function components obtained F With residuals R Together, they serve as multi-scale interval features. In the bivariate empirical mode decomposition process, the upper boundary sequence and the lower boundary sequence participate in the decomposition synchronously as two variables of the same interval sequence, so that the eigenmode function components of the same order correspond to the upper boundary component and the lower boundary component respectively, in order to maintain the phase correlation during the change process of the upper and lower boundaries.

[0065] in, and The input window at time step The upper and lower boundaries, This represents the total number of intrinsic mode function components. and The first The components corresponding to the upper and lower boundaries among the eigenmode function components, and These are the residual terms corresponding to the upper and lower boundaries, respectively.

[0066] Multi-scale graph spatiotemporal networks include graph convolutional long short-term memory network layers. GC - LSTM Attention layer that works in conjunction with spatiotemporal coordination. GC - LSTM The first layer aggregates spatial neighbor features and extracts long-term temporal dependencies based on the region association graph; the spatiotemporal collaborative attention layer employs a multi-head self-attention mechanism. ul - HeadA en on Adaptive allocation of spatiotemporal weights.

[0067] Before the input graph is convolved into a long short-term memory network layer, the intrinsic mode function components, residual terms, and enhanced environmental features corresponding to each region at the same time step are concatenated by region node to form the input features for the current time step. The graph convolutional long short-term memory network layer first performs graph convolution processing on the input features at the current time step based on the region association graph to obtain spatial aggregation features:

[0068] in, This is the normalized adjacency matrix of the region association graph. The input features at the current time step are obtained by fusing the multi-scale interval features and the two-layer environmental factors. and For learnable parameters, For activation function, This represents the spatial aggregation feature obtained after graph convolution processing.

[0069] The graph convolutional long short-term memory network layer then performs gated state updates based on the spatial aggregation features:

[0070]

[0071]

[0072]

[0073]

[0074]

[0075] in, , , These are the input gate, forget gate, and output gate, respectively. The current memory state, Currently in a hidden state. This is an element-wise multiplication operation.

[0076] Furthermore, the spatiotemporal collaborative attention layer employs a multi-head self-attention mechanism to weight the hidden states of environmental monitoring data units in different regions at different time points. The calculation method is as follows:

[0077] in:

[0078] in, r Indicates the first r One's attention, h To focus on the number of heads, , , and These are learnable parameters. Through the aforementioned multi-head self-attention mechanism, the spatiotemporal collaborative attention layer adaptively determines the contribution weights of each regional environmental monitoring data unit at different time points.

[0079] A multi-objective loss function is used to calculate the inference error and backpropagate to update the network weights. This multi-objective loss function includes numerical error loss, directional consistency loss, and interval consistency loss. Specifically, the numerical error loss constrains the upper and lower bounds of the inference to be close to the true upper and lower bounds of the sample labels, respectively; the directional consistency loss constrains the direction of change of the center value of the inference interval to be consistent with the direction of change of the center value of the true interval corresponding to the sample label; and the interval consistency loss constrains the upper bound of the inference to be no lower than the lower bound.

[0080] The center value of the extrapolated interval is the average of the extrapolated upper and lower boundaries, and the center value of the true interval is the average of the true upper and lower boundaries. Joint optimization using a multi-objective loss function enables the multi-scale graph-spatiotemporal network to simultaneously constrain interval boundary errors, the direction of interval center change, and interval boundary consistency.

[0081] The multi-objective loss function is expressed as:

[0082] The numerical error loss, which considers both the upper and lower bounds of the derivation, is expressed as:

[0083] The directional consistency loss is used to constrain the direction of change of the center value of the extrapolated interval to be consistent with the direction of change of the center value of the actual interval, and is expressed as:

[0084] in:

[0085]

[0086] in, For time step The true interval center value, For time step The central value of the derivation interval.

[0087] The interval consistency loss is used to constrain the upper boundary of the deduction to be no lower than the lower boundary of the deduction, and is expressed as:

[0088] in, For numerical error loss, For loss of directional consistency, For interval consistency loss, , and For the corresponding loss weight coefficient, and They are time steps The true upper boundary and the true lower boundary, and They are time steps The upper and lower bounds of the derivation, The true center value of the interval. To deduce the center value of the interval, N This is the length of the training window.

[0089] step S 6: Arrange the historical target interval window before the period to be simulated according to the steps. S 5. Process the data using the same decomposition method to obtain the current multi-scale interval features; update or reuse the regional correlation map based on the real target interval sequence before the time period to be simulated to obtain the current regional correlation map; obtain the environmental monitoring data corresponding to the time period to be simulated, obtain the daily environmental features corresponding to the time period to be simulated based on the environmental monitoring data, calculate the time-varying environmental pressure index corresponding to the time period to be simulated based on the daily environmental features, adjust the weight of the daily environmental features using the time-varying environmental pressure index, and generate the current two-layer environmental factor.

[0090] The current multi-scale interval features, the current regional correlation map, and the current two-layer environmental factors are input into the trained multi-scale graph spatiotemporal network. The multi-scale graph spatiotemporal network outputs the upper and lower boundaries of the time period to be inferred, and the numerical range formed by these two boundaries is used as the interval inference result.

[0091] Furthermore, after obtaining the interval extrapolation result, the target interval data subsequently collected is compared with the interval extrapolation result. When the actual collected target interval data exceeds the upper boundary of the extrapolation or falls below the lower boundary of the extrapolation, an abnormal state identifier is generated; when multiple consecutive time points generate abnormal state identifiers, a prompt is triggered to review the original data of the corresponding regional environmental monitoring data unit, re-collect environmental monitoring data, or provide incremental training prompts for the model. Thus, the interval extrapolation result not only serves as the output result of the target interval variation range but also as the input basis for subsequent abnormal state identification, data quality review, and model update processing.

[0092] This embodiment constructs a regional correlation graph through dynamic time warping distance, enabling the temporal misalignment relationships between environmental monitoring data units in different regions to participate in subsequent model training; it couples the time-varying environmental pressure index with daily environmental features through a two-layer environmental factor, improving the dynamic adaptability of environmental feature input; it generates extreme state target interval sequences through a conditional generative adversarial network containing standardized text semantic vectors and time-varying environmental pressure indices, helping to alleviate the problem of insufficient extreme state samples; and through joint training of multi-scale decomposition and multi-scale graph spatiotemporal networks, it can output interval inference results consisting of the inference upper boundary and the inference lower boundary, thereby providing data support for subsequent interval analysis.

[0093] like Figures 4-5 As shown, three regional environmental monitoring data units A , B , C right PM The 2.5-day average concentration range was used to extrapolate the regional distribution. A ,area B and region C Forming multiple consecutive historical time points PM 2.5 concentration range, unit is μ g / ³.

[0094] area A The historical target interval sequence formed within seven consecutive days can be: [32, 41], [36, 48], [43, 58], [51, 69], [48, 65], [40, 55], [35, 49]; area B The corresponding historical target interval sequences can be: [35, 46], [39, 52], [45, 61], [49, 67], [52, 71], [46, 63], [41, 56]; area CThe corresponding historical target interval sequence can be: [21, 30], [24, 34], [26, 38], [30, 42], [34, 47], [31, 44], [27, 39].

[0095] The above seven-day data is an example segment from a continuous historical target interval sequence. During model training, a continuous historical target interval sequence with a length sufficient to be divided into sliding windows according to a preset input window length and a preset inference step size is used.

[0096] At the same time, obtain the temperature, humidity, wind speed, and precipitation data for the corresponding time period in the above-mentioned areas. PM 10 concentration NO 2 concentration, SO 2 concentrations and O At least one of the three concentrations is used as a daily environmental feature, and time alignment, missing value imputation, outlier handling and normalization are performed to obtain the true target interval sequence and daily environmental features.

[0097] Using a seven-day period as the statistical cycle, if the extreme weather event frequency index is 0.30, the proportion of days exceeding air quality standards is 0.43, and the comprehensive pollutant concentration index is 0.62, with weights of 0.30, 0.30, and 0.40 respectively, then the time-varying environmental pressure index is 0.467. This time-varying environmental pressure index is input into the environmental gating unit, and the daily environmental characteristics are weighted to obtain enhanced environmental characteristics. The time-varying environmental pressure index and the enhanced environmental characteristics together constitute a two-layer environmental factor.

[0098] Based on region A ,area B and region C Calculate the dynamic time-warped distance of the true target interval sequence. For example, region A With the region B The dynamic time-normalized distance between them is 8.20, and the region... A With the region C The dynamic time-normalized distance between them is 31.60, and the region... B With the region C The dynamic time-warped distance between them is 28.40. When the distance scale parameter is set to 20, the corresponding region is obtained. A With the region B The edge weight between them is 0.664, and the region... A With the region C The edge weight between them is 0.206, and the region... B With the region C The edge weight between them is 0.242. The region... A ,area B and region CAs nodes, the edge weights are used as elements of the adjacency matrix to construct a regional association graph.

[0099] At least one of the following is used as the data processing specification text: pollutant monitoring limit standard, abnormal data judgment rule, monitoring equipment calibration rule, and data quality control rule. This text is then semantically encoded to obtain a specification text semantic vector. This semantic vector is fused with a time-varying environmental pressure index to form a conditional vector, which is then input into a conditional generative adversarial network (GAN) to generate a sequence of extreme state target intervals. For example, under high environmental pressure, the generated extreme state target interval sequence may include [62, 81], [66, 86], and [70, 92]. These three intervals are exemplary segments in the generated results. When constructing extreme state training samples, the GAN generates a continuous extreme state target interval sequence with a length sufficient to be divided into sliding windows according to a preset input window length and a preset inference step size.

[0100] Using the same preset input window length and preset inference step size, sliding windows are applied to both the real target interval sequence and the extreme state target interval sequence to obtain real training samples and extreme state training samples. Corresponding sample labels are then constructed to form an enhanced training sample set. Bivariate empirical mode decomposition is performed on the input window of each training sample. The resulting multi-scale interval features, region correlation maps, and corresponding two-layer environmental factors are input into a multi-scale graph spatiotemporal network, and joint optimization training is performed using the sample labels as supervised information.

[0101] During the extrapolation phase, the current multi-scale interval features, the current regional correlation map, and the current two-layer environmental factors are input into the trained multi-scale graph spatiotemporal network to obtain the regional... A Three time points in the future PM 2.5 Concentration projection range. For example, the output results are [38, 54] for day 8, [42, 61] for day 9, and [47, 68] for day 10.

[0102] If the area actually collected later A The target intervals are [40, 52] for day 8, [45, 64] for day 9, and [51, 73] for day 10. If the upper boundary of the actual collection intervals for day 9 and day 10 exceeds the upper boundary of the projection, the processor will generate an abnormal status indicator and trigger the original collection data verification, monitoring equipment status check, sampling data re-acquisition, or model incremental training prompt of the corresponding environmental monitoring data unit.

[0103] like Figures 6-7 As shown, the carbon emission range extrapolation for multiple functional areas within an industrial park is illustrated as an example. Multiple regional environmental monitoring data units include data units corresponding to the production workshop area. D Data units corresponding to the power station area Eand the data units corresponding to the warehousing and logistics areas F Each data unit is used to collect or generate at least one of the following: electricity consumption, gas consumption, equipment operating load, carbon emission accounting data, pollutant concentration data, and meteorological environmental data.

[0104] area D The carbon emission ranges formed within seven consecutive days can be: [118, 132], [125, 142], [139, 158], [146, 171], [152, 180], [141, 164], [130, 151]; area E The corresponding carbon emission ranges are: [92, 106], [98, 113], [110, 129], [121, 145], [132, 156], [124, 148], [116, 136]; area F The corresponding carbon emission ranges are: [35, 44], [38, 47], [41, 52], [45, 58], [49, 64], [47, 61], [42, 55], in units of... CO 2 e .

[0105] The above seven-day data is an example segment from a continuous historical target interval sequence. During model training, a continuous historical target interval sequence with a length sufficient to be divided into sliding windows according to a preset input window length and a preset inference step size is used.

[0106] At the same time, it acquires temperature, humidity, wind speed, precipitation, PM 2.5 concentration O 3. At least one of the following: concentration, production load, equipment operating load, electricity consumption, and gas consumption, shall be used as daily environmental characteristics or auxiliary operating data.

[0107] Using a seven-day period as the statistical cycle, if the extreme weather event frequency index is 0.57, the proportion of days exceeding air quality standards is 0.29, and the comprehensive pollutant concentration index is 0.52, with weights of 0.35, 0.25, and 0.40 respectively, then the time-varying environmental pressure index is 0.480. This time-varying environmental pressure index is input into the environmental gating unit, and the weights of temperature, pollutant concentration, production load, and equipment operating load are adjusted to obtain a two-layer environmental factor.

[0108] Based on region D ,area E and region F Calculate the dynamic time-warped distance of the true target interval sequence. For example, region D With the regionE The dynamic time-normalized distance between them is 12.50, and the region... D With the region F The dynamic time-normalized distance between them is 42.00, and the region... E With the region F The dynamic time-warped distance between them is 36.80. When the distance scale parameter is set to 25, the corresponding region is obtained. D With the region E The edge weight between them is 0.607, and the region... D With the region F The edge weight between them is 0.186, and the region... E With the region F The edge weights between them are 0.229, thus constructing a regional correlation graph for extrapolating carbon emission ranges.

[0109] At least one of the following is used as the data processing specification text: carbon emission accounting technical specifications, energy metering calibration rules, outlier data removal rules, missing energy consumption data compensation rules, and data quality control rules. This text is then semantically encoded to obtain a specification text semantic vector. This semantic vector is fused with a time-varying environmental pressure index to form a conditional vector, which is then input into a conditional generative adversarial network (GAN) to generate a sequence of extreme state target intervals under high load or high temperature disturbance conditions. For example, the generated region... D Extreme state carbon emission ranges may include [160, 194], [168, 205], and [174, 216]. These ranges are exemplary segments in the generated results; when actually constructing extreme state training samples, a continuous sequence of extreme state target ranges is generated.

[0110] Real training samples and extreme state training samples are constructed using the same preset input window length and preset inference step size, and sample labels are constructed for each training sample to obtain an enhanced training sample set. Bivariate empirical mode decomposition is performed on the input window of each training sample, and the resulting multi-scale interval features, regional correlation maps, and corresponding two-layer environmental factors are input into a multi-scale graph spatiotemporal network, using sample labels as supervision information for joint optimization training.

[0111] During the projection phase, the historical carbon emission range window from day 1 to day 7 was divided into steps. S 5. Using the same decomposition method, the current multi-scale interval features are obtained. These features, along with the current regional correlation graph and the current two-layer environmental factors, are then input into the trained multi-scale graph spatiotemporal network to obtain the regional... D The projected carbon emission ranges for the next three time points. For example, the output results are [136, 159] for day 8, [150, 178] for day 9, and [157, 188] for day 10.

[0112] If subsequent actual data collection or calculation yields the area D The carbon emission intervals are [138, 156] for day 8, [154, 181] for day 9, and [164, 201] for day 10. If the upper boundary of the actual carbon emission intervals for day 9 and day 10 exceeds the upper boundary of the projection, the processor generates an abnormal status flag and triggers the original energy consumption data review, energy metering equipment status check, emission factor configuration parameter verification, sampling data re-acquisition, or model incremental training prompt.

[0113] The above values ​​are only used to illustrate the data processing flow of the present invention and do not represent actual monitoring results, nor do they constitute a limitation on the number of regions, statistical period, monitoring indicators, time-varying environmental pressure index, gating weight, dynamic time warping distance, regional correlation edge weight, input window length, inference step size, or model output results.

[0114] like Figure 8 As shown, a time-series interval extrapolation system for multi-regional environmental monitoring data coupling includes a data processing module, a graph and factor construction module, an adversarial generation module, a sample construction module, a network training module, and an extrapolation module.

[0115] The data processing module is used to acquire historical target interval sequences, environmental monitoring data, and data processing specification texts from multiple regional environmental monitoring data units and perform preprocessing to obtain the actual target interval sequences and daily environmental characteristics, respectively.

[0116] The graph and factor construction module is used to calculate the dynamic time-normalized distance based on the real target interval sequence to construct the regional association graph; it calculates the time-varying environmental pressure index based on daily environmental characteristics, and uses the time-varying environmental pressure index to adjust the weight of daily environmental characteristics to generate a two-layer environmental factor.

[0117] The adversarial generative module is used to semantically encode the data processing standard text to obtain the standard text semantic vector. The standard text semantic vector is fused with the time-varying environmental pressure index to form a conditional vector. The conditional vector is then input into the conditional generative adversarial network for adversarial training to generate the extreme state target interval sequence.

[0118] The sample construction module is used to divide the real target interval sequence and the extreme state target interval sequence into sliding windows according to the preset input window length and the preset inference step size, respectively, to obtain real training samples and extreme state training samples; the target interval sequence corresponding to the preset inference step size after the input window of each training sample is used as the sample label to obtain the enhanced training sample set.

[0119] The network training module is used to decompose the input window of each training sample in the augmented training sample set to obtain multi-scale interval features. The multi-scale interval features, the region association map, and the two-layer environmental factors corresponding to the input window are input into the initial multi-scale graph spatiotemporal network. The sample labels are used as supervision information for joint optimization training to obtain the trained multi-scale graph spatiotemporal network.

[0120] The inference module is used to input the current multi-scale interval features, the current regional correlation map, and the current two-layer environmental factors corresponding to the time period to be inferred into the trained multi-scale graph spatiotemporal network for forward inference, and output the interval inference result composed of the upper and lower boundaries of the inference.

[0121] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements all the steps of the above-described time-series interval extrapolation method for coupling multi-regional environmental monitoring data.

[0122] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Equivalent substitutions or improvements made by those skilled in the art based on the disclosed technical solutions of this invention should all be included within the scope of protection of this invention.

Claims

1. A time-series interval extrapolation method for coupling multi-regional environmental monitoring data, characterized in that, Includes the following steps: step S 1. Obtain historical target interval sequences, environmental monitoring data, and data processing specification texts from multiple regional environmental monitoring data units and perform preprocessing to obtain the actual target interval sequences and daily environmental characteristics, respectively; wherein, the historical target interval sequences are interval-based operational status data collected or generated by regional environmental monitoring data units; step S 2: Calculate the dynamic time-warped distance based on the real target interval sequence to construct a regional correlation map; calculate the time-varying environmental pressure index based on the daily environmental characteristics, and use the time-varying environmental pressure index to adjust the weight of the daily environmental characteristics to generate a two-layer environmental factor; The steps S In step 2, the dynamic time-warped distance is calculated based on the real target interval sequence to construct a regional association graph, specifically including: Calculate the first The region and the first The dynamic time-warped distance between the real target interval sequences of each region is used to obtain a distance matrix; wherein, the real target interval sequence is a two-dimensional interval vector sequence composed of upper and lower boundary values, and the dynamic time-warped distance is calculated based on the two-dimensional interval vector sequence; The calculation of the dynamic time warping distance includes: determining the local distance between binary interval vectors based on the difference between the upper and lower boundaries of two regions at different time points, and determining the dynamic time warping distance based on the local distance according to the minimum cumulative distance of the dynamic time warping path; The distance matrix is ​​converted into edge weights using a distance scale parameter; Using multiple regional environmental monitoring data units as nodes and the edge weights as elements of an adjacency matrix, an undirected weighted graph is constructed as the regional association graph. The formula for calculating the edge weights is as follows: in, For the first The region and the first Edge weights between regions For the first The region and the first Dynamic time-normalized distance between regions For distance scale parameters, and ; The steps S In step 2, a time-varying environmental pressure index is calculated based on the daily environmental characteristics, and the weights of the daily environmental characteristics are adjusted using the time-varying environmental pressure index to generate a two-layer environmental factor, specifically including: The time-varying environmental pressure index is obtained by weighted summing of the extreme weather event frequency index, the proportion of days exceeding air quality standards, and the comprehensive pollutant concentration index. The calculation formula is as follows: in, For the first The time-varying environmental stress index for each statistical period. For the first The frequency index of extreme weather events over a statistical period For the first The percentage of days with air quality exceeding standards for each statistical period. For the first The comprehensive index of pollutant concentrations for each statistical period. , and For the corresponding component weights, and , and All are not less than 0. ; The time-varying environmental pressure index is input into the environmental gating unit to generate gating weights, and the gating weights are then multiplied element-wise with the daily environmental characteristics to obtain enhanced environmental characteristics. The time-varying environmental pressure index and the enhanced environmental characteristics together constitute the two-layer environmental factor, and the calculation formula for the enhanced environmental characteristics is as follows: in, For the first The gating weights corresponding to each daily time point For activation function, and For learnable parameters, Indicates the first The statistical period to which each daily time point belongs. For the first The time-varying environmental stress index corresponding to the statistical period of each daily time point. For the first Daily environmental characteristics at a specific time of day. For the first Enhanced environmental characteristics at individual daily time points This is an element-wise multiplication operation; the gate weights are scalars or vectors with the same dimension as the daily environmental features. step S 3: Semantically encode the data processing standard text to obtain a standard text semantic vector. Fuse the standard text semantic vector with the time-varying environmental pressure index to form a condition vector. Input the condition vector into a conditional generative adversarial network for adversarial training to generate an extreme state target interval sequence. step S 4. Divide the real target interval sequence according to the input window length and the inference step size to construct real training samples and their sample labels; divide the extreme state target interval sequence according to the same input window length and inference step size to construct extreme state training samples and their sample labels; mix the real training samples with the extreme state training samples to obtain an enhanced training sample set. step S 5: Decompose the input window of each training sample in the enhanced training sample set to obtain multi-scale interval features, and input the multi-scale interval features, the region association map and the two-layer environmental factors into the initial multi-scale graph spatiotemporal network. Use the sample labels as supervision information to perform joint optimization training to obtain the trained multi-scale graph spatiotemporal network. step S 6: Input the current multi-scale interval features, the current regional association map, and the current two-layer environmental factors corresponding to the time period to be inferred into the trained multi-scale graph spatiotemporal network for forward inference to obtain the interval inference result including the upper and lower boundaries of the inference.

2. The time-series interval extrapolation method for coupling multi-regional environmental monitoring data according to claim 1, characterized in that, The conditional generative adversarial network includes a generator and a discriminator; The steps S 3 also includes: Obtain historical data processing specification text, perform word segmentation, stop word removal, and vectorization on the historical data processing specification text, and use word vector models, pre-trained language models, or... ransfor er The encoding model extracts semantic features to obtain the semantic vector of the standard text; The conditional vector is obtained by concatenating the standard text semantic vector with the time-varying environmental pressure index. The conditional vector and random noise are input into the generator, which outputs the corresponding target interval sequence. The generated target interval sequence and the condition vector are input together into the discriminator, and the real target interval sequence and the condition vector are input together into the discriminator, which then performs an authenticity evaluation. After the generator and the discriminator are trained adversarially until the changes in both the generator loss and the discriminator loss are less than a preset loss threshold, or after the number of training rounds reaches a preset number of rounds, the generator output target interval sequence is extracted. The generated target interval sequence is subjected to boundary consistency correction to ensure that the upper boundary of generation is not lower than the lower boundary of generation. Then, the generated target interval sequence with time-varying environmental pressure index greater than a preset pressure threshold or target interval fluctuation amplitude greater than a preset fluctuation threshold is selected from the generated target interval sequence after boundary consistency correction to obtain the extreme state target interval sequence.

3. The time-series interval extrapolation method for coupling multi-regional environmental monitoring data according to claim 1, characterized in that, The steps S 4. Specifically includes: According to the preset input window length and inference step size, the real target interval sequence is divided into multiple real training samples. Each real training sample includes an input window consisting of real target intervals at multiple consecutive time points, and a sample label located after the input window as supervision information. The extreme state target interval sequence is divided into multiple extreme state training samples according to the same input window length and inference step size. Each extreme state training sample includes an input window consisting of extreme state target intervals at multiple consecutive time points, and a sample label located after the input window as supervision information. The real training samples and the extreme state training samples are mixed according to a preset mixing ratio to obtain the enhanced training sample set.

4. The time-series interval extrapolation method for coupling multi-regional environmental monitoring data according to claim 1, characterized in that, The steps S In step 5, the input window of each training sample in the enhanced training sample set is decomposed to obtain multi-scale interval features, specifically including: Bivariate empirical mode decomposition is used to jointly decompose the upper and lower boundary sequences in the input window, separating multiple intrinsic mode function components and a residual term respectively; All the intrinsic mode function components and all the residual terms are used together as the multi-scale interval features; The calculation formula for the joint decomposition is: in, and The input window at time step The upper and lower boundaries, This represents the total number of intrinsic mode function components. and The first The components corresponding to the upper and lower boundaries among the eigenmode function components, and These are the residual terms corresponding to the upper and lower boundaries, respectively.

5. The time-series interval extrapolation method for coupling multi-regional environmental monitoring data according to claim 1, characterized in that, The multi-scale graph spatiotemporal network includes a graph convolutional long short-term memory network layer and a spatiotemporal collaborative attention layer connected in sequence. The steps S 5 specifically includes: Using the graph convolutional long short-term memory network layer based on the region association graph, spatial neighbor feature aggregation and long-term temporal dependency extraction are performed on the multi-scale interval features and the two-layer environmental factors to obtain basic spatiotemporal features. The multi-head self-attention mechanism in the spatiotemporal collaborative attention layer is used to adaptively allocate the weights of environmental monitoring data units in each region at different time points, and the basic spatiotemporal features are weighted and mapped into higher-order spatiotemporal features. Based on the higher-order spatiotemporal features, the upper and lower bounds of the inference are output, and the inference error is calculated according to the upper bound, the lower bound and the corresponding sample labels. The network weights of the multi-scale graph spatiotemporal network are then updated by backpropagation.

6. The time-series interval extrapolation method for coupling multi-regional environmental monitoring data according to claim 5, characterized in that, The joint optimization training uses a multi-objective loss function to calculate the inference error, which includes numerical error loss, directional consistency loss, and interval consistency loss. The numerical error loss is used to constrain the upper and lower boundaries of the inference to be close to the upper and lower boundaries of the labels in the sample labels, respectively. The directional consistency loss is used to constrain the direction of change of the center value of the extrapolation interval to be consistent with the direction of change of the center value of the label interval corresponding to the sample label; The interval consistency loss is used to constrain the upper boundary of the deduction to be no lower than the lower boundary of the deduction; the interval consistency loss includes a boundary penalty term generated when the lower boundary of the deduction is higher than the upper boundary of the deduction. Wherein, the center value of the deduction interval is the average of the upper and lower boundaries of the deduction, and the center value of the label interval is the average of the upper and lower boundaries of the label.

7. The time-series interval extrapolation method for coupling multi-regional environmental monitoring data according to claim 1, characterized in that, The steps S 6 specifically includes: The historical target interval window before the period to be simulated is arranged according to the steps. S 5. Process the data using the same decomposition method to obtain the current multi-scale interval features; The current regional association map is obtained by updating or reusing the actual target interval sequence before the time period to be simulated. Obtain environmental monitoring data corresponding to the time period to be simulated, obtain daily environmental characteristics corresponding to the time period to be simulated based on the environmental monitoring data, calculate the time-varying environmental pressure index corresponding to the time period to be simulated based on the daily environmental characteristics, adjust the weight of the daily environmental characteristics using the time-varying environmental pressure index, and generate the current two-layer environmental factor. The current multi-scale interval features, the current region association map, and the current two-layer environmental factors are jointly input into the trained multi-scale graph spatiotemporal network. The multi-scale graph spatiotemporal network, after training, outputs the upper and lower boundaries of the time period to be inferred, and uses the numerical range formed by the upper and lower boundaries as the interval inference result.

8. A time-series interval extrapolation system for coupling multi-regional environmental monitoring data, characterized in that, The time-series interval extrapolation method for performing multi-regional environmental monitoring data coupling according to any one of claims 1 to 7 includes: The data processing module is configured to acquire historical target interval sequences, environmental monitoring data, and data processing specification texts from multiple regional environmental monitoring data units and perform preprocessing to obtain the actual target interval sequences and daily environmental characteristics, respectively. The graph and factor construction module is configured to calculate the dynamic time-warped distance based on the real target interval sequence to construct a regional association graph, calculate the time-varying environmental pressure index based on the daily environmental characteristics, and use the time-varying environmental pressure index to adjust the weight of the daily environmental characteristics to generate a two-layer environmental factor. The adversarial generation module is configured to semantically encode the data processing standard text to obtain a standard text semantic vector, fuse the standard text semantic vector with the time-varying environmental pressure index to form a conditional vector, and input the conditional vector into a conditional generative adversarial network for adversarial training to generate an extreme state target interval sequence. The sample construction module is configured to divide the real target interval sequence and the extreme state target interval sequence according to the input window length and the inference step size, respectively, to construct real training samples, extreme state training samples and corresponding sample labels, and to mix the real training samples and the extreme state training samples to obtain an enhanced training sample set. The network training module is configured to decompose the input window of each training sample in the enhanced training sample set to obtain multi-scale interval features, and input the multi-scale interval features, the region association graph and the two-layer environmental factors into the initial multi-scale graph spatiotemporal network, and perform joint optimization training with the sample labels as supervision information to obtain the trained multi-scale graph spatiotemporal network. The inference module is configured to input the current multi-scale interval features, the current regional association map, and the current two-layer environmental factors corresponding to the time period to be inferred into the trained multi-scale graph spatiotemporal network for forward inference, and output the interval inference result including the upper and lower boundaries of the inference.

Images