Surface temperature prediction method and system based on multilayer perception adaptation and prompt guidance

Abstract

The invention relates to a surface temperature prediction method and system based on multi-layer perception adaptation and prompt guidance, belongs to the technical field of space-time prediction, solves the problem of low prediction accuracy caused by the fact that inherent heterogeneity in space-time data is not considered in the prior art, and comprises the following steps: S1, obtaining multi-scale time priori knowledge; collecting historical spatio-temporal data in the spatial region to be predicted and processing the historical spatio-temporal data to obtain a standard data set; s2, constructing a prediction model based on multi-layer perception adaptation and prompt collaborative guidance, and generating space-time representation after space-time prompt guidance; s3, adopting the standard data set to train the constructed prediction model based on the multi-layer perception adaptation and prompt collaborative guidance to obtain a trained model; and S4, predicting future spatio-temporal data to obtain spatio-temporal representation after generating spatio-temporal prompt guidance, taking the spatio-temporal representation as a spatio-temporal prediction result, and providing the spatio-temporal prediction result to a surface air temperature monitoring process.

Description

Technical Field

[0001] The present invention relates to the technical field of spatiotemporal data mining, and in particular to a surface air temperature prediction method and system based on multi-layer perception adaptation and prompt guidance. Background Art

[0002] Providing surface air temperature forecasts is used to prejudge abnormal temperatures and issue early warnings, as well as to deploy protection against extreme cold and extreme high temperature weather in advance. For example, it plays an important role and significance in protecting crops from abnormal temperatures, allocating electricity demand in urban and rural areas to abnormal high temperatures, and preventing forest fires to abnormal high temperatures, etc., to the normal and safe operation of urban and rural production and life.

[0003] The goal of spatiotemporal forecasting is to predict future trends in specific spatiotemporal dimensions by analyzing and modeling historical spatiotemporal data. Accurate future trend predictions can provide a scientific basis for various decision-making processes, promoting the optimization of social management and resource allocation. Therefore, spatiotemporal forecasting has important application value in the field of surface temperature monitoring and has attracted widespread attention.

[0004] Existing spatiotemporal prediction methods used in surface air temperature monitoring typically model the spatial characteristics of spatiotemporal data as a graph structure, where nodes and edges represent spatial units and their interdependencies, respectively. Temporal characteristics are modeled as time series of nodes on the graph, using spatiotemporal graph networks to capture spatiotemporal correlations. However, these methods often focus too much on global average characteristics, assuming that pattern variations are uniform across all spatial regions or time periods. This ignores the inherent heterogeneity in spatiotemporal data, namely, the differences and unevenness in the distribution and patterns of data across different spatial regions or time periods.

[0005] This field requires an improved spatiotemporal prediction method that can provide efficient, accurate and comprehensive spatiotemporal prediction for complex scenarios with high spatiotemporal heterogeneity in applications such as surface temperature monitoring and early warning and temperature anomaly protection layout. Summary of the Invention

[0006] In view of the above problems, the present invention provides a surface air temperature prediction method and system based on multi-layer perception adaptation and prompt guidance, which solves the problem of ignoring temporal and spatial heterogeneity in the prior art.

[0007] According to one embodiment of the present invention, a surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance is provided, comprising the following steps:

[0008] Step S1, obtaining multi-scale time prior knowledge, and collecting historical spatiotemporal data in the spatial area to be predicted and performing standardization processing to obtain a standard data set including historical spatiotemporal sequences, wherein the collected historical spatiotemporal data includes surface temperature data;

[0009] Step S2, constructing a prediction model based on multi-layer perception adaptation and prompt collaborative guidance, wherein the prediction model includes a spatiotemporal prompt collaborative generation module and a spatiotemporal prompt guidance module, which is used to generate a spatiotemporal representation after spatiotemporal prompt guidance;

[0010] Step S3, using the standard data set to train the constructed prediction model based on multi-layer perception adaptation and prompt collaborative guidance to obtain a trained model;

[0011] In step S4, the trained prediction model is applied to the prediction of future spatiotemporal data to obtain the spatiotemporal representation after generating spatiotemporal prompt guidance, which is used as the spatiotemporal prediction result of the surface temperature. The spatiotemporal prediction result is provided to the surface temperature monitoring process for temperature warning and climate anomaly protection monitoring configuration planning and layout.

[0012] Optionally, in step S1: standardizing the historical spatiotemporal data within the spatial area to be predicted includes dividing the spatial area to be predicted into N sub-areas, counting the historical spatiotemporal data of surface air temperature monitoring in each time period in each sub-area, and inputting the historical spatiotemporal data of each time period in each sub-area into a three-dimensional array in the format of B*T*N to obtain a standard data set including historical spatiotemporal sequences, wherein B represents the data batch, T represents the time series length, and N represents the total number of spatial nodes. Each data point in the standard data set represents the surface air temperature of the node in the data batch at a time step; the multi-scale time prior knowledge includes: month, date, time step and week.

[0013] Optionally, step S2 specifically includes the following steps:

[0014] Step S2.1: Build the backbone network of the prediction model, which captures spatiotemporal dependencies and generates spatiotemporal representations through stacked graph convolutional networks and temporal convolutional networks.

[0015] Step S2.2: construct a spatiotemporal cue collaborative generation module to generate specific cue vectors in time and space to capture spatiotemporal heterogeneity;

[0016] Step S2.3, construct a spatiotemporal cue guidance module, input the fused spatiotemporal cues into the multi-layer perception adapter, flexibly map them to the corresponding feature space, and generate a spatiotemporal representation guided by spatiotemporal cues.

[0017] Optionally, step S2.2 specifically includes the following steps:

[0018] Step S2.2.1, construct the temporal cue generation part, map the multi-scale temporal prior knowledge from discrete to continuous feature space, obtain the mapped temporal prior knowledge, and based on the mapped temporal prior knowledge, adaptively learn the importance of different prior knowledge through the feature fusion network to obtain the multi-scale temporal fusion representation E t , as a time reminder;

[0019] Step S2.2.2, build the spatial cue generation part and randomly initialize a trainable node embedding for each spatial node To capture spatial heterogeneity, the trainable node embedding E s As a spatial reminder, Indicates domain, d h The feature dimension representing the multi-scale temporal fusion representation;

[0020] Step S2.2.3, construct a dynamic gate fusion module, combining the obtained time hint E t With space prompt E s , perform dimension expansion, feature splicing and fusion to obtain the fused spatiotemporal prompts.

[0021] Optionally, step S2.2.1 specifically includes the following steps:

[0022] First, the multi-scale time prior knowledge is mapped from discrete to continuous feature space to obtain the mapped time prior knowledge:

[0023] e i =Embedding(t i )

[0024] Among them, e i represents the embedding vector of the i-th time prior knowledge, as the time prior knowledge after mapping, Embedding(·) represents the embedding operation, t i represents the value of the i-th time prior knowledge, i represents the number of the time prior knowledge and i=1,2,…,M, M is the total number of selected time prior knowledge;

[0025] Then, based on the mapped temporal prior knowledge, the importance of different prior knowledge is adaptively learned through the feature fusion network to obtain a multi-scale temporal fusion representation:

[0026] E t =concat(e1,e2,…e i ,…,e M )W t +b t

[0027] Among them, E tIt is a multi-scale temporal fusion representation and T represents the total length of the time step, concat(·) represents the concatenation process, and W t is the weight of the feature fusion network and sum(·) represents the sum function, d i represents the embedding dimension of the i-th time prior knowledge, b t is the bias of the feature fusion network and

[0028] The obtained multi-scale time fusion representation E t As a time reminder.

[0029] Optionally, step S2.2.3 specifically includes the following steps:

[0030] Construct a dynamic gating fusion module and combine it with the generated time hint E t With space prompt E s , perform dimension expansion and feature splicing, and calculate the gating weight:

[0031]

[0032] Where G is the gating weight and σ(·) is the Sigmoid function, which ensures that the gate weight is in the interval [0,1], W g represents the weight of gated fusion, b g represents the bias of gated fusion, Represents a splicing operation.

[0033] Next, the temporal and spatial cues are dynamically gated and fused using the obtained gating weights:

[0034] E st =G⊙E t +(1-G)⊙E s

[0035] Among them, E st represents the fused spatiotemporal cues and ⊙ represents the Hadamard product operation.

[0036] Optionally, step S2.3 specifically includes the following steps:

[0037] In step S2.3.1, the encoders and decoders of each layer of the prediction model process the historical spatiotemporal data and spatiotemporal representations to obtain the spatiotemporal representations of each layer before the spatiotemporal cue guidance;

[0038] The fused spatiotemporal cues are feature transformed by the perceptual adapter layer corresponding to the encoder and decoder layers of the multi-layer perceptual adapter to obtain the spatiotemporal cues of this layer:

[0039]

[0040] Among them, ReLU(·) represents the ReLU activation function, They represent the two linear mapping weights of the l-th layer perception adapter, l represents the network layer number of the perception adapter, MLP (l) (E st ) represents the spatiotemporal cue transformed by the layer l perception adapter;

[0041] In step S2.3.2, the spatiotemporal cue transformed by the multi-layer perceptual adapter is subjected to a Hadamard product and residual connection with the spatiotemporal representation obtained by the encoder and decoder of this layer to obtain the spatiotemporal representation guided by the spatiotemporal cue:

[0042] H (l)′ =H (l) +H (l) MLP (l) (E st )

[0043] Among them, H (t) is the first layer of spatiotemporal representation before spatiotemporal cue guidance, and H (l)′ It is the first layer of spatiotemporal representation after the guidance of spatiotemporal cues;

[0044] The obtained spatiotemporal representation guided by the spatiotemporal cues is processed by the decoder to obtain and output the predicted value of the future spatiotemporal sequence.

[0045] Optionally, step S3 includes obtaining the mean absolute error as the prediction loss:

[0046]

[0047] Among them, L mae Represents the prediction loss, n represents the number of spatial nodes, q represents the time step within the prediction window, represents the predicted value of the future space-time series, Represents the true value of the future space-time series;

[0048] Based on the obtained prediction loss, the gradient is calculated through back propagation, and the parameters of the prediction model are updated using the stochastic gradient descent algorithm to obtain a trained prediction model.

[0049] According to another embodiment of the present invention, a surface air temperature prediction system based on multi-layer perception adaptation and prompt guidance is provided, including:

[0050] The data collection and processing module is used to collect historical spatiotemporal data in the spatial area to be predicted, standardize the historical spatiotemporal data, and obtain a standard data set;

[0051] The prediction model based on multi-layer perceptual adaptation and prompt collaborative guidance is a unified framework with an encoder and decoder. It encodes historical spatiotemporal sequences through graph convolutional networks and temporal convolutional networks to capture spatial and temporal dependencies, captures spatiotemporal heterogeneity by generating spatiotemporal cues, and then guides the spatiotemporal modeling process to overcome the interference of spatiotemporal heterogeneity.

[0052] The training module uses a standard dataset to train the prediction model.

[0053] Optionally, the prediction model based on multi-layer perception adaptation and prompt collaborative guidance includes:

[0054] The encoder and decoder capture spatiotemporal dependencies and generate spatiotemporal representations layer by layer through stacked graph convolutional networks and temporal convolutional networks;

[0055] The spatiotemporal cue collaborative generation module generates specific cue vectors in time and space to capture spatiotemporal heterogeneity, and balances the contributions of spatiotemporal cues through a dynamic gating fusion mechanism to collaboratively generate comprehensive spatiotemporal cues.

[0056] The spatiotemporal cue guidance module inputs spatiotemporal cues into the perception adapters at each layer and flexibly maps them to the corresponding feature space to obtain the spatiotemporal representation guided by spatiotemporal cues;

[0057] The spatiotemporal prompt collaborative generation module includes:

[0058] The temporal cue generation part maps prior knowledge from discrete to continuous, generates temporal cue vectors through a feature fusion network, and enables the prediction model to perceive the differences between different temporal contexts.

[0059] The spatial cue generation part initializes an adaptive node embedding for each spatial node as a spatial cue vector, enabling the prediction model to perceive the differences between different spatial contexts;

[0060] The dynamic gated fusion module balances the contributions of spatiotemporal cues through a dynamic gated fusion mechanism to collaboratively generate comprehensive spatiotemporal cues.

[0061] Compared with the prior art, the surface air temperature prediction method and system based on multi-layer perception adaptation and prompt guidance provided by the present invention have at least the following beneficial effects.

[0062] 1) Capture spatiotemporal heterogeneity through flexible spatiotemporal cue generation, balance spatiotemporal cue contributions and perform cue fusion using a dynamic gated fusion mechanism, map global spatiotemporal cues to the feature space of each layer using a multi-layer perception adapter to adapt to feature representations at different levels, and efficiently guide the spatiotemporal modeling process using spatiotemporal cues, thereby overcoming the interference of spatiotemporal heterogeneity in surface temperature monitoring and providing accurate and robust predictions in surface temperature monitoring, which is conducive to achieving more optimized and accurate temperature warnings, as well as climate anomaly protection and layout.

[0063] 2) Through the time cue generation part, the prior knowledge is mapped from discrete to continuous, and the time cue vector is generated through the feature fusion network, so that the prediction model can perceive the differences between different time contexts and take into account the impact of time factors on surface temperature, such as seasonal and day-night changes, to improve the accuracy of the prediction.

[0064] 3) Through the spatial cue generation part, an adaptive node embedding is initialized for each spatial node as a spatial cue vector, which enables the prediction model to perceive the differences between different spatial contexts and take into account the temperature differences in different geographical locations, such as the impact of factors such as terrain and latitude on surface temperature, to improve the accuracy of the prediction.

[0065] 4) Set up a dynamic gated fusion module to balance the contribution of spatiotemporal cues through the dynamic gated fusion mechanism, collaboratively generate comprehensive spatiotemporal cues, comprehensively consider the joint effect of temporal and spatial factors, and provide more comprehensive and accurate information for prediction.

[0066] 5) Set up a spatiotemporal cue guidance module, input the spatiotemporal cues into the perception adapters of each layer, flexibly map them to the corresponding feature space, obtain the spatiotemporal representation after spatiotemporal cue guidance, further optimize the spatiotemporal characteristics, make the prediction model better adapt to spatiotemporal heterogeneity, and improve the accuracy of surface temperature prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. By referring to the drawings, the features and advantages of the present invention can be more clearly understood. The drawings are schematic and should not be understood as limiting the present invention in any way. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0068] Figure 1 A schematic diagram of a surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance according to an embodiment of the present invention.

[0069] Figure 2A schematic block diagram of a surface air temperature prediction system based on multi-layer perception adaptation and prompt guidance provided according to an embodiment of the present invention.

[0070] Figure 3 It is a part of the historical spatiotemporal data collected in the spatial area to be predicted in an embodiment of the surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance provided according to an embodiment of the present invention.

[0071] Figure 4 This is a curve comparison chart of the surface air temperature prediction value and the actual value obtained by applying an embodiment of the surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present invention and the features therein can be combined with each other without conflict.

[0073] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0074] The following describes in detail a surface air temperature prediction method and system based on multi-layer perception adaptation and prompt guidance according to an embodiment of the present invention with reference to the accompanying drawings.

[0075] As shown in the figure, the surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance provided in accordance with an embodiment of the present invention includes the following steps.

[0076] like Figure 1 As shown, a surface air temperature prediction method and system based on multi-layer perception adaptation and prompt guidance provided according to one embodiment of the present invention is used for spatiotemporal prediction of surface air temperature monitoring, including first collecting historical spatiotemporal data in the spatial area to be predicted, and constructing a standard data set based on this; then constructing a prediction model based on multi-layer perception adaptation and prompt collaborative guidance, the prediction model including a spatiotemporal prompt collaborative generation module and a spatiotemporal prompt guidance module; using a standard data set to complete the training of the prediction model; finally, applying the trained prediction model to the prediction of future spatiotemporal data.

[0077] According to an embodiment of the present invention, a surface air temperature prediction method and system based on multi-layer perception adaptation and prompt guidance are provided, which are used for spatiotemporal prediction of surface air temperature monitoring and include the following steps.

[0078] Step S1: Acquire multi-scale temporal prior knowledge, collect historical spatiotemporal data within the spatial region to be predicted, and perform standardization processing to obtain a standard dataset. The collected historical spatiotemporal data may be historical spatiotemporal data from surface air temperature monitoring. Standardizing the historical spatiotemporal data within the spatial region to be predicted includes dividing the spatial region to be predicted into N subregions, compiling historical spatiotemporal data from surface air temperature monitoring for each time period within each subregion, and inputting the historical spatiotemporal data for each time period within each subregion into a three-dimensional array in the format of B*T*N to obtain a standard dataset comprising a historical spatiotemporal sequence, where B represents the data batch, T represents the time series length, and N represents the total number of spatial nodes. Each data point in the standard dataset represents the surface air temperature at that node within the data batch at a time step. The acquired historical spatiotemporal data is used as a historical spatiotemporal sequence. The aforementioned temporal prior knowledge corresponds one-to-one with the time dimension of the historical spatiotemporal data. The multi-scale temporal prior knowledge may include information such as month, day, time step, weekday, whether it is a weekday, and whether it is a holiday, obtained from a standard calendar.

[0079] Step S2: constructing a prediction model based on multi-layer perceptual adaptation and spatiotemporal cue collaborative guidance, wherein the prediction model includes a spatiotemporal cue collaborative generation module and a spatiotemporal cue guidance module.

[0080] This prediction model based on multi-layer perceptual adaptation and prompt collaborative guidance is a unified framework with an encoder and a decoder. It encodes historical spatiotemporal sequences through graph convolutional networks and temporal convolutional networks to capture spatial and temporal dependencies, captures spatiotemporal heterogeneity by generating spatiotemporal prompts, and then guides the spatiotemporal modeling process to overcome the interference of spatiotemporal heterogeneity.

[0081] The prediction model based on multi-layer perception adaptation and prompt collaborative guidance is as follows:

[0082]

[0083] in, is a prediction model based on multi-layer perceptual adaptation and prompt collaborative guidance, θ represents all trainable parameters in the prediction model, X t-P+1:t represents the historical spatiotemporal data from t-P+1 to time t, represents the future space-time data from t+1 to t+q, X represents the historical real space-time sequence, Represents the future prediction spatiotemporal sequence, t represents the time step, P represents the historical window size, and Q represents the future window size.

[0084] Step S2 specifically includes the following steps.

[0085] Step S2.1: Build the backbone network of the prediction model. The encoder and decoder capture spatiotemporal dependencies and generate spatiotemporal representations through stacked graph convolutional networks and temporal convolutional networks. Taking layer l as an example, the graph convolution operation and temporal convolution operation are defined as:

[0086]

[0087] Among them, Z (l) is the l-th layer of spatiotemporal representation. When l = 0, it is the spatiotemporal sequence input. is the hidden representation obtained by graph convolution, W (l) is the mapping weight, σ(·) is the GELU activation function, Conv1D2(·) and Conv1D1(·) are two layers of one-dimensional convolution operations, A adp is the adaptive adjacency matrix.

[0088] Adaptive adjacency matrix powered by trainable node embeddings To build:

[0089]

[0090] in, represents the domain, d represents the feature dimension of multi-scale temporal fusion representation, and N is the total number of spatial nodes.

[0091] Step S2.2 constructs a spatiotemporal cue collaborative generation module to generate specific cue vectors in both time and space to capture spatiotemporal heterogeneity. Temporally, multi-scale temporal prior knowledge is mapped from discrete to continuous, and a feature fusion network is used to generate temporal cue vectors, enabling the prediction model to perceive the differences between different temporal contexts. Spatially, adaptive node embeddings are initialized for each spatial node as spatial cue vectors, enabling the prediction model to perceive the differences between different spatial contexts. Finally, a dynamic gated fusion mechanism is used to balance the contributions of spatiotemporal cues, collaboratively generating comprehensive spatiotemporal cues. This step S2.2 specifically includes the following steps.

[0092] Step S2.2.1, constructing a time prompt generation part for generating time prompts.

[0093] First, the multi-scale time prior knowledge is mapped from discrete to continuous feature space to obtain the mapped time prior knowledge:

[0094] e i =Embedding(t i )

[0095] Among them, e i represents the embedding vector of the i-th time prior knowledge, Embedding represents the embedding operation, t iRepresents the value of the i-th time prior knowledge, i represents the number of the time prior knowledge, and i=1,2,…,M, M is the total number of selected time prior knowledge.

[0096] Then, based on the mapped temporal prior knowledge, the importance of different prior knowledge is adaptively learned through the feature fusion network to obtain a multi-scale temporal fusion representation:

[0097] E t =concat(e1,e2,…e i ,…,e M )W t +b t

[0098] Among them, E t It is a multi-scale temporal fusion representation and W t is the weight of the feature fusion network and b t is the bias of the feature fusion network and T represents the total length of the time step, sum(·) represents the summation function, d i represents the embedding dimension of the i-th time prior knowledge, and concat(·) represents the concatenation process.

[0099] The obtained multi-scale temporal fusion representation E t , i.e., the generated multi-scale temporal cues, which capture temporal heterogeneity.

[0100] In addition, unlike spatiotemporal sequences, temporal prior knowledge is known before prediction. Therefore, when making spatiotemporal predictions, corresponding time cues are generated for both historical windows and future windows, thereby comprehensively guiding the spatiotemporal modeling process.

[0101] By constructing the time cue generation part above, the prior knowledge is mapped from discrete to continuous, and the time cue vector is generated through the feature fusion network, so that the prediction model can perceive the differences between different time contexts and take into account the impact of time factors on surface temperature, such as seasonal and day-night changes, to improve the accuracy of the prediction.

[0102] Step S2.2.2, build the spatial cue generation part and randomly initialize a trainable node embedding for each spatial node To capture spatial heterogeneity, the trainable node embeddings serve as spatial cues.

[0103] By constructing a spatial cue generation part, an adaptive node embedding is initialized for each spatial node as a spatial cue vector, which enables the prediction model to perceive the differences between different spatial contexts and take into account the temperature differences in different geographical locations, such as the impact of factors such as terrain and latitude on surface temperature, thereby improving the accuracy of the prediction.

[0104] Step S2.2.3, construct a dynamic gate fusion module, combining the time prompt E generated in the above steps t With space prompt E s , perform dimension expansion and feature splicing, and calculate the gate weight Dynamic balance of space-time prompts contribution:

[0105]

[0106] Among them, W g represents the weight of gated fusion, b g represents the bias of gated fusion, represents the concatenation operation, σ(·) is the Sigmoid function, which ensures that the gating weight is in the interval [0,1].

[0107] Next, the temporal and spatial cues are dynamically gated and fused using the obtained gating weights to obtain the fused spatiotemporal cues:

[0108] E st =G⊙E t +(1-G)⊙E s

[0109] Where ⊙ represents the Hadamard product operation, E st represents the fused spatiotemporal cues and It captures comprehensive spatial and temporal heterogeneity.

[0110] By constructing the dynamic gating fusion module as described above, the contribution of spatiotemporal cues is balanced by the dynamic gating fusion mechanism, comprehensive spatiotemporal cues are generated collaboratively, and the joint effects of temporal and spatial factors are comprehensively considered to provide more comprehensive and accurate information for prediction.

[0111] In step S2.3, a spatiotemporal cue guidance module is constructed, and the fused spatiotemporal cues are input into the multi-layer perception adapter, flexibly mapped to the corresponding feature space, and the spatiotemporal representation guided by the spatiotemporal cues is obtained, thereby guiding the spatiotemporal modeling process of the prediction model to overcome the interference of heterogeneity on spatiotemporal modeling.

[0112] Specifically, the spatiotemporal cue guidance module first transforms the spatiotemporal cues through a multi-layer perception adapter, and then accurately guides the spatiotemporal representation of each layer of the prediction model, thereby overcoming the interference of spatiotemporal heterogeneity, improving prediction accuracy and robustness, and making the prediction model more adaptable to dynamic and complex spatiotemporal changes. Step S2.3 specifically includes the following steps.

[0113] Step S2.3.1: The encoders and decoders of each layer of the prediction model based on multi-layer perceptual adaptation and prompt collaborative guidance process the historical spatiotemporal data and spatiotemporal representation to obtain the spatiotemporal representation H of each layer before spatiotemporal prompt guidance. (l) , l represents the number of layers of encoder and decoder.

[0114] The fused spatiotemporal cues are feature transformed by the perceptual adapter layer corresponding to the encoder and decoder of this layer in the multi-layer perceptual adapter to obtain spatiotemporal cues suitable for the feature space of this layer:

[0115]

[0116] Among them, ReLU(·) represents the ReLU activation function, Represents the twice linear mapping weights of the l-th layer perception adapter, MLP (l) (·) represents the lth layer of the multi-layer perception adapter, l represents the network layer number of the perception adapter, MLP (l) (E st ) represents the spatiotemporal cues transformed by the layer l perception adapter.

[0117] In step S2.3.2, the spatiotemporal cue transformed by the multi-layer perceptual adapter is subjected to a Hadamard product and residual connection with the spatiotemporal representation obtained by the encoder and decoder of this layer to obtain the spatiotemporal representation guided by the spatiotemporal cue:

[0118] H (l)′ =H (l) +H (l) ⊙MLP (l) (E st )

[0119] Among them, H (l) is the first layer of spatiotemporal representation before spatiotemporal cue guidance, and H (l)′ It is the lth layer of spatiotemporal representation after the guidance of spatiotemporal cues.

[0120] The obtained spatiotemporal representation guided by the spatiotemporal cues is processed by the decoder to obtain and output the predicted value of the future spatiotemporal sequence.

[0121] By constructing the spatiotemporal cue guidance module as described above, the spatiotemporal cues are input into the perception adapters at each layer and flexibly mapped to the corresponding feature space to obtain the spatiotemporal representation after spatiotemporal cue guidance. The spatiotemporal features are further optimized, so that the prediction model can better adapt to spatiotemporal heterogeneity and improve the accuracy of surface air temperature prediction.

[0122] Step S3: Use the standard data set to train the prediction model to obtain a trained model.

[0123] The mean absolute error (MAE) is used as the prediction loss to optimize the prediction model, as shown in the following formula:

[0124]

[0125] Among them, L mae represents the prediction loss, n represents the number of spatial nodes and n=1,2…N, q represents the time step in the prediction window, Q represents the future window size, represents the predicted value of the future space-time series, Represents the true value of the future space-time series.

[0126] Based on the obtained prediction loss, the gradient is calculated through back propagation, and the parameters of the prediction model are updated using the stochastic gradient descent algorithm to obtain a trained prediction model.

[0127] Step S4: Apply the trained prediction model to future spatiotemporal data to obtain spatiotemporal surface temperature prediction results. These spatiotemporal prediction results can be provided to the surface temperature monitoring process for temperature warnings and climate anomaly protection planning and layout.

[0128] The following references Figure 2 A surface air temperature prediction system based on multi-layer perception adaptation and prompt guidance provided according to another embodiment of the present invention is described.

[0129] like Figure 2 As shown, a surface air temperature prediction system based on multi-layer perception adaptation and prompt guidance provided according to another embodiment of the present invention includes: a data collection and processing module, which is used to collect historical spatiotemporal data in the spatial area to be predicted, standardize the historical spatiotemporal data, and obtain a standard data set, and the acquired historical spatiotemporal data include: surface air temperature; a prediction model based on multi-layer perception adaptation and prompt collaborative guidance is a unified framework with an encoder and a decoder, which encodes historical spatiotemporal sequences through graph convolutional networks and time convolutional networks to capture spatial and temporal dependencies, captures spatiotemporal heterogeneity by generating spatiotemporal prompts, and then guides the spatiotemporal modeling process to overcome the interference of spatiotemporal heterogeneity; a training module, which uses a standard data set to train the prediction model.

[0130] The prediction model based on multi-layer perception adaptation and prompt collaborative guidance may include: an encoder and a decoder, which capture spatiotemporal dependencies and generate spatiotemporal representations layer by layer through stacked graph convolutional networks and temporal convolutional networks; a spatiotemporal prompt collaborative generation module, which generates specific prompt vectors in time and space to capture spatiotemporal heterogeneity, and balances the contribution of spatiotemporal prompts through a dynamic gating fusion mechanism to collaboratively generate comprehensive spatiotemporal prompts; a spatiotemporal prompt guidance module, which inputs spatiotemporal prompts into each layer of perception adapters, flexibly maps them to the corresponding feature space, and obtains spatiotemporal representations after spatiotemporal prompt guidance.

[0131] The spatiotemporal cue collaborative generation module may include: a temporal cue generation part, which maps prior knowledge from discrete to continuous, generates a temporal cue vector through a feature fusion network, and enables the prediction model to perceive the differences between different temporal contexts; a spatial cue generation part, which initializes an adaptive node embedding for each spatial node as a spatial cue vector, and enables the prediction model to perceive the differences between different spatial contexts; a dynamic gating fusion module, which balances the contribution of spatiotemporal cues through a dynamic gating fusion mechanism, and collaboratively generates comprehensive spatiotemporal cues.

[0132] The spatiotemporal cue guidance module may include a multi-layer perception adapter for flexibly mapping the fused spatiotemporal cues to the corresponding feature space to generate a spatiotemporal representation guided by the spatiotemporal cue.

[0133] Example 1

[0134] For ease of understanding, the surface air temperature pressure in a city is used as an example to provide a more detailed exemplary description of the surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance of the present invention.

[0135] Step S1 collects statistical data on Beijing's surface air temperature and pressure, collected by temperature sensors at N spatial locations over a period of time. The surface air temperature data is then standardized, with the surface air temperature at each spatial location and time period counted as a standard dataset. The standard dataset is a three-dimensional array of size B × T × N, where B is the number of batches, T is the length of the time series, and N is the number of spatial locations.

[0136] Step S2: construct a prediction model based on multi-layer perception adaptation and prompt collaborative guidance (STMP2G).

[0137] Step S3: training the prediction model using the standard data set.

[0138] Step S4: input the historical spatiotemporal data of the spatial region to be predicted into the trained prediction model, and output the future spatiotemporal data of the surface air temperature of each region.

[0139] In this embodiment 1, the Beijing surface air temperature data set is used to train the prediction model. The statistical data from January 1, 2015 to December 31, 2018 are intercepted, covering a total of 184 spatial locations, with a time interval of 3 hours. The data set is divided into a training set (70%), a validation set (10%), and a test set (20%) in the time dimension. In this embodiment, the surface air temperature of the past week is used to predict the surface air temperature of the next week. The training set is used to train the model, the validation set is used to save the optimal model, and the test set is used to test the model performance. Figure 3This is a portion of the historical spatiotemporal data collected within the spatial region to be predicted in Example 1. As an example, only surface temperature data for six spatial locations from midnight to midnight on January 1, 2015, in Kelvin (K) is displayed. The constructed deep learning model is trained using the resulting training set. The data is Z-score normalized, and all parameters in the prediction model based on multi-layer perceptual adaptation and prompt collaborative guidance are randomly initialized.

[0140] Training was performed on an Intel(R) Xeon(R) Processors CPU and an NVIDIA Tesla V100S GPU on a Linux operating system using the PyTorch framework, with the batch size set to 64 and the initial learning rate set to 0.001.

[0141] Using AdamW as the optimizer, this deep learning model was trained for 200 epochs on the complete training set. The model was validated using a loss function at each training epoch, and the optimal model was saved based on the loss function value. An early stopping strategy was used during training, terminating training early if the loss function value did not decrease for 20 consecutive epochs. Figure 4 The curve comparison chart of the surface air temperature prediction value obtained in Example 1 and the actual value shows that the fitting result of the prediction value and the actual value is better, providing a more accurate prediction result.

[0142] The prediction results of the above embodiment were compared with those of the prior art. Surface air temperature prediction was performed on the same dataset. The comparison results are shown in Table 1. The prediction results were evaluated using mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). Lower errors indicate better prediction results. Example 1 was compared with nine prior art prediction methods.

[0143] The first is STGCN, which integrates graph convolution and gated temporal convolution via spatiotemporal convolution blocks to capture spatiotemporal dependencies.

[0144] The second is GraphWaveNet, which proposes an adaptive adjacency matrix to capture the hidden spatial dependencies in the data and captures the temporal dependencies through stacked dilated one-dimensional convolutions.

[0145] The third is AGCRN, which captures node-specific patterns through node-adaptive parameter learning, captures potential spatial dependencies with unified node embeddings, and automatically captures spatiotemporal correlations in combination with recurrent networks.

[0146] The fourth is DMSTGCN, which extends the existing graph convolution to dynamic graph convolution by constructing an adaptive spatial adjacency matrix with a day-cycle to model the changing spatial correlation.

[0147] The fifth is STSSL, which adaptively enhances traffic flow graph data at the data level and topological structure level, and captures spatiotemporal heterogeneity through auxiliary self-supervised learning tasks.

[0148] The sixth is MegaCRN, which captures spatiotemporal heterogeneity through spatiotemporal meta-graph learning and combines graph convolution and recurrent networks on the meta-graph to capture spatiotemporal correlation.

[0149] The seventh is TESTAM, which selects a specific spatial modeling method for each node through a hybrid expert model and combines it with a time-enhanced attention model to capture spatiotemporal dependencies.

[0150] The eighth is TGCRN, which learns a time-varying graph structure that can perceive periodicity and trends, and combines it with a gated recurrent unit to jointly capture dynamic spatiotemporal dependencies.

[0151] The ninth is HimNet, which implicitly captures spatiotemporal heterogeneity through spatiotemporal embedding and learns spatiotemporal specific parameters from a meta-parameter pool.

[0152] It can be clearly seen from the comparison results in Table 1 that the surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance in Example 1 has better prediction results than the nine prediction methods in the prior art.

[0153] Table 1 Comparison of surface temperature prediction results of Beijing dataset

[0154] Comparison plan MAE(K) RMSE(K) MAPE (%) STGCN 3.4857 4.1470 1.1652 GraphWaveNet 3.3524 4.0610 1.1780 AGCRN 3.6996 4.2338 1.3345 DMSTGCN 3.4904 4.1838 1.2672 STSSL 3.5668 4.2178 1.2801 MegaCRN 3.4525 4.0986 1.2693 TESTAM 3.8240 4.5043 1.3642 TGCRN 3.8573 4.3833 1.3884 HimNet 3.5779 4.2330 1.2858 Example 1 3.1484 3.9371 1.1039

[0155] Furthermore, several ablation experiments were conducted to demonstrate the effectiveness of the prediction method of the present invention. The experimental results are shown in Table 2.

[0156] Example 1 was compared with five post-ablation models.

[0157] (1) basemodel: The spatiotemporal cue guidance module is deleted, representing the base model that only captures spatiotemporal dependencies but cannot capture spatiotemporal heterogeneity.

[0158] (2) w / o Et: The time prompt generation module was deleted.

[0159] (3) w / o Es: The spatial hint generation module was deleted.

[0160] (4) w / o G: The spatiotemporal cue fusion module was deleted.

[0161] (5) w / o Adapter: The layer-by-layer adapter is removed, and the same spatiotemporal cues are used to guide the spatiotemporal representation of each layer.

[0162] Table 2 Comparison of ablation experiment results on Beijing dataset

[0163] Comparison plan MAE(K) RMSE(K) MAPE (%) basemodel 3.2360 4.0438 1.1321 w / o Et 3.2232 4.0089 1.1236 w / o Es 3.2187 4.0310 1.1214 w / o G 3.1813 3.9580 1.1226 w / o Adapter 3.2290 3.9838 1.1458 STMP2G (Example 1) 3.1484 3.9371 1.1039

[0164] It can be seen from the experimental results in Table 2 that the prediction effect of the prediction method of Example 1 of the present invention is better than that of the five post-ablation models.

[0165] All of the above optional technical solutions can be combined in any way to form optional embodiments of the present application, and will not be described in detail here.

[0166] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0167] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance, characterized in that: The following steps are involved: Step S1, obtaining multi-scale time prior knowledge, and collecting historical spatiotemporal data in the spatial area to be predicted and performing standardization processing to obtain a standard data set including historical spatiotemporal sequences, wherein the collected historical spatiotemporal data includes surface temperature data; Step S2, constructing a prediction model based on multi-layer perception adaptation and prompt collaborative guidance, wherein the prediction model includes a spatiotemporal prompt collaborative generation module and a spatiotemporal prompt guidance module, which is used to generate a spatiotemporal representation after spatiotemporal prompt guidance; Step S3, using the standard data set to train the constructed prediction model based on multi-layer perception adaptation and prompt collaborative guidance to obtain a trained model; In step S4, the trained prediction model is applied to the prediction of future spatiotemporal data to obtain the spatiotemporal representation after generating spatiotemporal prompt guidance, which is used as the spatiotemporal prediction result of surface temperature and airflow. The spatiotemporal prediction result is provided to the surface temperature monitoring process for temperature warning and climate anomaly protection planning and layout.

2. The surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance according to claim 1 is characterized in that: In step S1: Standardizing the historical spatiotemporal data within the spatial region to be predicted includes dividing the spatial region to be predicted into N subregions, counting the historical spatiotemporal data of surface air temperature monitoring within each subregion for each time period, and inputting the historical spatiotemporal data within each subregion for each time period into a three-dimensional array in the format of B*T*N to obtain a standard data set including historical spatiotemporal sequences, where B represents a data batch, T represents a time series length, and N represents the total number of spatial nodes. Each data point in the standard data set represents the surface air temperature of the node within the data batch at a time step. The multi-scale time prior knowledge includes: month, date, time step and week.

3. The surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance according to claim 1 is characterized in that: Step S2 specifically includes the following steps: Step S2.1: Build the backbone network of the prediction model, which captures spatiotemporal dependencies and generates spatiotemporal representations through stacked graph convolutional networks and temporal convolutional networks. Step S2.2: construct a spatiotemporal cue collaborative generation module to generate specific cue vectors in time and space to capture spatiotemporal heterogeneity; Step S2.3, construct a spatiotemporal cue guidance module, input the fused spatiotemporal cues into the multi-layer perception adapter, flexibly map them to the corresponding feature space, and generate a spatiotemporal representation guided by spatiotemporal cues.

4. The surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance according to claim 3 is characterized in that: Step S2.2 specifically includes the following steps: Step S2.2.1, construct the temporal cue generation part, map the multi-scale temporal prior knowledge from discrete to continuous feature space, obtain the mapped temporal prior knowledge, and based on the mapped temporal prior knowledge, adaptively learn the importance of different prior knowledge through the feature fusion network to obtain the multi-scale temporal fusion representation E t , as a time reminder; Step S2.2.2, build the spatial cue generation part and randomly initialize a trainable node embedding for each spatial node To capture spatial heterogeneity, the trainable node embedding E s As a spatial reminder, Indicates domain, d h The feature dimension representing the multi-scale temporal fusion representation; Step S2.2.3, construct a dynamic gate fusion module, combining the obtained time hint E t With space prompt E s , perform dimension expansion, feature splicing and fusion to obtain the fused spatiotemporal prompts.

5. The surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance according to claim 4 is characterized in that: Step S2.2.1 specifically includes the following steps: First, the multi-scale time prior knowledge is mapped from discrete to continuous feature space to obtain the mapped time prior knowledge: e i =Embedding(t i ) Among them, e i represents the embedding vector of the i-th time prior knowledge, as the time prior knowledge after mapping, Embedding(·) represents the embedding operation, t i represents the value of the i-th time prior knowledge, i represents the number of the time prior knowledge and i=1,2,…,M, M is the total number of selected time prior knowledge; Then, based on the mapped temporal prior knowledge, the importance of different prior knowledge is adaptively learned through the feature fusion network to obtain a multi-scale temporal fusion representation: AND t =concat(e1,e2,…e i ,…,and M )W t +b t Among them, E t It is a multi-scale temporal fusion representation and T represents the total length of the time step, concat(·) represents the concatenation process, and W t is the weight of the feature fusion network and sum(·) represents the sum function, d i represents the embedding dimension of the i-th time prior knowledge, b t is the bias of the feature fusion network and The obtained multi-scale time fusion representation E t As a time reminder.

6. The surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance according to claim 5 is characterized in that: Step S2.2.3 specifically includes the following steps: Construct a dynamic gating fusion module and combine it with the generated time hint E t With space prompt E s , perform dimension expansion and feature splicing, and calculate the gating weight: Where G is the gating weight and σ(·) is the Sigmoid function, which ensures that the gate weight is in the range [0,1], W g represents the weight of gated fusion, b g represents the bias of gated fusion, Represents a splicing operation; Next, the temporal and spatial cues are dynamically gated and fused using the obtained gating weights: E st =G⊙E t +(1-G)⊙E s Among them, E st represents the fused spatiotemporal cues and ⊙ represents the Hadamard product operation.

7. The surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance according to claim 6 is characterized in that: Step S2.3 specifically includes the following steps: In step S2.3.1, the encoders and decoders of each layer of the prediction model process the historical spatiotemporal data and spatiotemporal representations to obtain the spatiotemporal representations of each layer before the spatiotemporal cue guidance; The fused spatiotemporal cues are feature transformed by the perceptual adapter layer corresponding to the encoder and decoder layers of the multi-layer perceptual adapter to obtain the spatiotemporal cues of this layer: Among them, ReLU(·) represents the ReLU activation function, They represent the two linear mapping weights of the l-th layer perception adapter, l represents the network layer number of the perception adapter, MLP (l) (E st ) represents the spatiotemporal cue transformed by the layer l perception adapter; In step S2.3.2, the spatiotemporal cue transformed by the multi-layer perceptual adapter is subjected to a Hadamard product and residual connection with the spatiotemporal representation obtained by the encoder and decoder of this layer to obtain the spatiotemporal representation guided by the spatiotemporal cue: H (l)′ =H (l) +H (l) ⊙MLP (l) (E st ) Among them, H (l) is the first layer of spatiotemporal representation before spatiotemporal cue guidance, and H (l)′ It is the first layer of spatiotemporal representation after the guidance of spatiotemporal cues; The obtained spatiotemporal representation guided by the spatiotemporal cues is processed by the decoder to obtain and output the predicted value of the future spatiotemporal sequence.

8. The surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance according to claim 7 is characterized in that: Step S3 involves obtaining the mean absolute error as the prediction loss: Among them, L mae Represents the prediction loss, n represents the number of spatial nodes, q represents the time step within the prediction window, represents the predicted value of the future space-time series, Represents the true value of the future space-time series; Based on the obtained prediction loss, the gradient is calculated through back propagation, and the parameters of the prediction model are updated using the stochastic gradient descent algorithm to obtain a trained prediction model.

9. A system for implementing the surface air temperature prediction method based on multi-layer perception adaptation and prompt guidance according to any one of claims 1 to 8, characterized in that: include: A data collection and processing module is used to collect historical spatiotemporal data in the spatial area to be predicted and perform standardization processing to obtain a standard data set including historical spatiotemporal sequences; The prediction model based on multi-layer perceptual adaptation and prompt collaborative guidance is a unified framework with an encoder and decoder. It encodes historical spatiotemporal sequences through graph convolutional networks and temporal convolutional networks to capture spatial and temporal dependencies, captures spatiotemporal heterogeneity by generating spatiotemporal cues, and then guides the spatiotemporal modeling process to overcome the interference of spatiotemporal heterogeneity. The training module uses a standard data set to train the prediction model.

10. The system according to claim 9, characterized in that The prediction model based on multi-layer perception adaptation and prompt collaborative guidance includes: The encoder and decoder capture spatiotemporal dependencies and generate spatiotemporal representations layer by layer through stacked graph convolutional networks and temporal convolutional networks; The spatiotemporal cue collaborative generation module generates specific cue vectors in time and space to capture spatiotemporal heterogeneity, and balances the contributions of spatiotemporal cues through a dynamic gating fusion mechanism to collaboratively generate comprehensive spatiotemporal cues. The spatiotemporal cue guidance module inputs spatiotemporal cues into the perception adapters at each layer and flexibly maps them to the corresponding feature space to obtain the spatiotemporal representation guided by spatiotemporal cues; The spatiotemporal prompt collaborative generation module includes: The temporal cue generation part maps prior knowledge from discrete to continuous, generates temporal cue vectors through a feature fusion network, and enables the prediction model to perceive the differences between different temporal contexts. The spatial cue generation part initializes an adaptive node embedding for each spatial node as a spatial cue vector, enabling the prediction model to perceive the differences between different spatial contexts; The dynamic gated fusion module balances the contributions of spatiotemporal cues through a dynamic gated fusion mechanism to collaboratively generate comprehensive spatiotemporal cues.

Images