Urban heat island intensity prediction method based on physical mechanism guided recurrent neural network

Abstract

The invention discloses an urban heat island intensity prediction method based on a physical mechanism guided recurrent neural network. The method comprises the following steps: S1, constructing a multi-source fusion data set including land utilization cover data and meteorological data; and S2, constructing a PGRNN fusion model fusing a recurrent neural network RNN and a long short-term memory network LSTM, inputting the urban heat island intensity sequence data or the multi-source fusion data set and the urban heat island intensity matching combination sequence data into the PGRNN fusion model to carry out prediction training, and constructing an urban heat island intensity prediction expression to carry out heat island intensity prediction processing. The PGRNN fusion model realizes long-time-sequence prediction training of a multi-source fusion data set and urban heat island intensity, a physical guidance mechanism taking urban canopy energy balance as a core is creatively introduced and constructed in the PGRNN fusion model, and an urban canopy energy balance physical rule plays a key constraint role in a prediction process of long-time-sequence data; and the consistency and the precision of the urban heat island intensity under a long time sequence are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of urban climate and environment monitoring, and particularly relates to a method for predicting urban heat island intensity based on a physical mechanism guided recurrent neural network. BACKGROUND

[0002] Urban heat island (UHI) is a typical urban climate phenomenon, which is characterized by the fact that the temperature in urban areas is significantly higher than that in surrounding areas. It is the most perceptible urban climate phenomenon for humans. Heat waves (HW) refer to consecutive high-temperature weather. With the intensification of global climate change, HW occurs frequently worldwide, leading to an increase in the incidence and mortality of related diseases and is considered one of the most deadly natural disasters. HW can cause increased energy consumption, reduced agricultural and pastoral production, and more conflict incidents in cities. The interaction between urban heat island effect and heat waves has a serious impact on urban environment and human health. On the one hand, the intensity of urban heat island is stronger during the heat wave period than during the non-heat wave period. On the other hand, the urban heat island effect increases the impact range, intensity and duration of heat waves.

[0003] However, due to the lack of high-spatial and temporal resolution urban heat island data, there are few studies on the intra-daily spatial variation characteristics of the HW-UHI effect. The current research direction of urban heat island data is to process and output results from meteorological monitoring data, remote sensing thermal infrared data and numerical simulation. However, there are the following technical problems: (1) meteorological monitoring data has limited spatial coverage, making it difficult to finely reflect the changes in the spatial form of the heat island, and the ability to capture local details of the urban heat island UHI is insufficient, which cannot meet the needs of fine research on the spatial effect of the heat island; (2) remote sensing thermal infrared data has low temporal resolution due to the revisit period of satellites, making it difficult to capture the hourly changes of the urban heat island UHI within a day, which hinders the study of the dynamic evolution process of the heat island effect in time; (3) numerical models usually have high computational complexity, especially under extreme high-temperature conditions, the simulation error is more significant, and the accuracy of input parameters is very high. If the parameters are slightly biased, it may affect the accuracy and reliability of the simulation results. SUMMARY

[0004] The purpose of the present application is to solve the technical problems pointed out in the background art, provide a city heat island intensity prediction method and system based on physical mechanism guided recurrent neural network, use multi-source fusion data set and city heat island intensity matching combination sequence data as the data input and prediction output of PGRNN fusion model, realize long time sequence prediction training of multi-source fusion data set and city heat island intensity in PGRNN fusion model, creatively introduce the construction of physical guiding mechanism with city canopy energy balance as the core in PGRNN fusion model, and the city canopy energy balance physical rule plays a key constraint role in the prediction process of long time sequence data, and improves the consistency and accuracy of city heat island intensity under long time sequence.

[0005] The purpose of the present application is realized by the following technical solutions:

[0006] A city heat island intensity prediction method based on physical mechanism guided recurrent neural network, the method comprises:

[0007] S1, constructing a multi-source fusion data set containing land use cover data and meteorological data;

[0008] S2, constructing a PGRNN fusion model combining recurrent neural network RNN and long short-term memory network LSTM, the PGRNN fusion model inputs city heat island intensity sequence data or multi-source fusion data set and city heat island intensity matching combination sequence data for prediction training and constructs the following city heat island intensity prediction expression:

[0009] , wherein , are city heat island intensities at time t and t-1 respectively, is a city area temperature change value at time t-1, is a city surrounding suburb temperature change value at time t-1.

[0010] In order to better realize the present application, the city area temperature change value at time t-1 and the city surrounding suburb temperature change value at time t-1 are obtained by the following formula:

[0011] , wherein is a net radiation change amount, is a ground heat storage net flux change amount, is a human heat flux change amount, is a proportional coefficient, is a correction coefficient, is an energy redistribution parameter, , is Stefan-Boltzmann constant, is the surface temperature.

[0012] Preferably, the PGRNN fusion model is constructed as follows:

[0013] ; wherein is the net shortwave radiation, is the net longwave radiation; is the sensible heat flux, is the latent heat flux, is the net ground heat flux.

[0014] Preferably, the change in the net ground heat flux is calculated and retrieved by the objective lag model using the multi-source fusion dataset according to the following expression:

[0015] , wherein , , are urban empirical parameters, respectively; is the net radiation, is the time;

[0016] is the anthropogenic heat flux change is the anthropogenic heat flux is calculated and retrieved by the following formula:

[0017] , wherein is the population density, is the total population.

[0018] Preferably, the net shortwave radiation is expressed as: ; the net longwave radiation is expressed as: , the sensible heat flux is expressed as: ; the latent heat flux is expressed as: ; wherein is the average surface albedo, is the shortwave radiation reaching the ground, , are the surface emissivity and the atmospheric emissivity, respectively, is the Stefan-Boltzmann constant, , are the air temperature and the surface temperature, respectively, is the average air density, is the specific heat of air at constant pressure, is the air dynamic resistance, is the urban heat island intensity.

[0019] Preferably, the time t-1 urban heat island intensity The expression is as follows:

[0020] , wherein is the temperature data of the urban area at time t-1 (the temperature data corresponds to the selection of the air temperature or the surface temperature of the urban area at time t-1), is the temperature data of the suburban area surrounding the city at time t-1 (the temperature data corresponds to the selection of the air temperature or the surface temperature of the suburban area surrounding the city at time t-1).

[0021] Preferably, the PGRNN fusion model is verified and constrained by comparing the temperature measured value of the meteorological station at the meteorological station site with the temperature prediction value output by the PGRNN fusion model at the meteorological station site and calculating the root mean square error.

[0022] Preferably, the joint loss constraint function expression of the PGRNN fusion model is as follows:

[0023] , is the mean square error of the output prediction value and the true value of the recurrent neural network RNN, or the mean square error calculated by comparing the temperature measured value of the meteorological station at the meteorological station site with the temperature prediction value output by the recurrent neural network RNN in the PGRNN fusion model at the meteorological station site; is the error between the prediction value and the true value of the urban heat island intensity output by the PGRNN fusion model, is the dynamically adjusted weight coefficient.

[0024] Preferably, in the method S1, the meteorological data sources include reanalysis meteorological data and meteorological station monitoring data; the land use and cover data includes multiple land cover type data sets classified according to urban areas and suburban areas surrounding the city.

[0025] A physical mechanism guided recurrent neural network based urban heat island intensity prediction system, comprising a multi-source fusion sample data set, a data acquisition module and a PGRNN fusion model, the multi-source fusion sample data set is a multi-source fusion data set matched and combined with urban heat island intensity sequence data or urban heat island intensity sequence data, the data acquisition module is used to acquire data containing land use and cover data and meteorological data in real time and construct a multi-source fusion data set; the PGRNN fusion model is constructed by fusing a recurrent neural network RNN and a long short-term memory network LSTM, and the PGRNN fusion model constructs the following urban heat island intensity prediction expression: , wherein , Urban heat island intensity at time t, t-1 respectively, Urban heat island intensity at time t-1 in the urban area, Urban heat island intensity at time t-1 in the surrounding suburb of the city; the PGRNN fusion model uses a multi-source fusion sample data set for model prediction training.

[0026] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0027] (1) The present application uses a multi-source fusion data set and a matching combination sequence data of urban heat island intensity as the data input and prediction output of the PGRNN fusion model, and the PGRNN fusion model realizes long-time sequence prediction training of the multi-source fusion data set and the urban heat island intensity. A physical guiding mechanism is creatively introduced in the PGRNN fusion model, which is centered on the urban canopy energy balance. The physical rules of urban canopy energy balance play a key constraint role in the prediction process of long-time sequence data, and improve the consistency and accuracy of urban heat island intensity under long-time sequence.

[0028] (2) The PGRNN fusion model of the present application introduces a physical guiding mechanism, which can output urban heat island intensity prediction results with a time resolution of hours and a spatial resolution of less than 1 km, realizing fine-grained dynamic monitoring of urban heat island intensity, and being suitable for heat environment prediction and early warning under different research cities and various climate conditions. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The present application is a method flow chart of the urban heat island intensity prediction method;

[0030] Figure 2 The present application is a processing principle diagram of the PGRNN fusion model in the embodiment;

[0031] Figure 3 The present application is a model pseudo code diagram of the PGRNN fusion model in the embodiment;

[0032] Figure 4 The present application is a distribution diagram of urban heat island intensity prediction of a certain research city in the research month and research time in the embodiment. DETAILED DESCRIPTION

[0033] The present application will be further described in detail below in combination with embodiments:

[0034] EMBODIMENT

[0035] As shown in the figure, a city heat island intensity prediction method based on a physical mechanism guided recurrent neural network, the method comprises: Figure 1

[0036] ​S1, construct a multi-source fusion dataset containing land use and cover data and meteorological data. In this embodiment, the meteorological data sources include reanalysis meteorological data and meteorological station monitoring data (including temperature, wind speed, precipitation, etc. monitoring data), and the reanalysis meteorological data can be derived from FNL (Final Operational Global Analysis) data, which is global reanalysis data provided by the National Environmental Prediction Center (NECP) / National Center for Atmospheric Research (NCAR). The land use and cover data includes a variety of land cover type data sets classified by urban area, urban peripheral suburb, and land cover types including arable land, forest land, impervious surface, wetland, bare land, grassland, water area, etc. The land use and cover data can be derived from the China Land Cover Dataset (CLCD), which has spatial distribution information of land use and cover.

[0037] In this embodiment, a WRF-UCM fusion model (Weather Research and Forecasting Model with Urban Canopy Model, i.e. Weather Research and Forecasting Model integrated with Urban Canopy Model) is constructed by fusing a WRF model (Weather Research and Forecasting, referred to as WRF) and an urban canopy model UCM. The WRF model uses 5 layers of nesting, with the center longitude and latitude being 120.16° E / 30.25° N, and the corresponding network resolution being 8.1 km (Domain1), 2.7 km (Domain2), 900 m (Domain3), 300 m (Domain4), and 100 m (Domain5), respectively. The vertical direction is divided into 27 layers, and the top of the model is about 20 km high. The layers are divided into sparse upper layers and dense lower layers to more accurately simulate the boundary layer structure, with about 8 layers below 1 km. The initial meteorological field boundary condition is 1°x 1° reanalysis meteorological data with a time resolution of 6 h. The physical parameter scheme of the WRF-UCM fusion model is selected as follows: Lin microphysical process scheme, Dudhia shortwave radiation scheme, RRTM longwave radiation scheme, MYJ boundary layer condition, Noah land surface process scheme, and UCM urban canopy scheme. The land use and cover data of the outermost layer (Domain 1) is the 30 s resolution data provided by the model, and the innermost layer (Domain 5) is the 30 m land use and cover data from the China Land Cover Dataset CLCD.

[0038] S2, construct a PGRNN fusion model of a recurrent neural network RNN and a long short-term memory network LSTM physical guidance mechanism (English full name: Physics-Guided Recurrent Neural Network), the PGRNN fusion model is constructed based on a recurrent neural network RNN and a long short-term memory network LSTM to construct an energy conservation-based physical mechanism (i.e. a physical guidance mechanism), the energy conservation physical mechanism adopts the following energy balance formula of the urban canopy: ; wherein is the net shortwave radiation, is the net longwave radiation; is the sensible heat flux, is the latent heat flux, is the net ground heat flux.

[0039] Among them, the net shortwave radiation is expressed as: , is the average ground reflectivity, is the shortwave radiation reaching the ground. The net longwave radiation is the difference between the downward longwave radiation and the upward longwave radiation , that is , the upward longwave radiation is expressed as , the downward longwave radiation is expressed as: , the net longwave radiation is expressed as: , , are the ground emissivity and atmospheric emissivity, is the Stefan-Boltzmann constant, , are the air temperature and the ground temperature. The sensible heat flux is expressed as: ; the latent heat flux is expressed as: ; is the average air density (the value taken in this embodiment is 1.21 kg / m 3 ), is the specific heat of air under constant pressure, the value taken in this embodiment is , is the aerodynamic resistance, is the Bowen ratio.

[0040] , , , is the energy redistribution parameter, is the net radiation; combining the above formula, we get: , and the derivative is: ; since is very small, after rearrangement we get:

[0041] , into the formula , rearranged as follows:

[0042] .

[0043] The long short-term memory network LSTM can utilize the urban heat island intensity sequence data or the multi-source fusion data set and the urban heat island intensity matching combination sequence data to input long time sequence data and output subsequent prediction; the PGRNN fusion model can utilize the urban heat island intensity sequence data to input long time sequence data and output subsequent prediction; as shown in Figure 2 , Figure 3 , the PGRNN fusion model mainly utilizes the multi-source fusion data set and the urban heat island intensity matching combination sequence data to input long time sequence data and output subsequent prediction; the multi-source fusion data set and the urban heat island intensity are input as independent variables and dependent variables into the PGRNN fusion model, that is, the multi-source fusion data set at a certain time point is input to realize the prediction output of the urban heat island intensity. As shown in Figure 2 , the PGRNN fusion model is fused with the recurrent neural network RNN and the long short-term memory network LSTM; the long short-term memory network LSTM has a gating mechanism (forget gate, input gate and output gate) and a cell state (long-term memory channel), and the recurrent unit of the recurrent neural network RNN is reconstructed, Figure 2 , wherein represents an input sequence, that is, a multi-source fusion data set arranged in time sequence; Figure 2 , wherein represents a hidden state, which is used to store historical information up to the current time step, at the same time, depending on the current input and the hidden state of the last moment, the historical information is transmitted to the current step through'recycling'; Figure 2 , wherein represents an output sequence, which is an urban heat island intensity sequence data (i.e. predicting future urban heat island intensity). The PGRNN fusion model inputs the urban heat island intensity sequence data or the multi-source fusion data set and the urban heat island intensity matching combination sequence data to perform prediction training and construct the following urban heat island intensity prediction expression:

[0044] , wherein , The urban heat island intensity at times t and t-1 are respectively. This represents the temperature change over time t-1 in the urban area. This represents the temperature change over time t-1 in the suburbs surrounding the city. For example... Figure 2 As shown, the PGRNN fusion model of this invention is constructed with a physical mechanism guided by the core principle of urban canopy energy balance. When the PGRNN fusion model drives prediction, the input from the previous time step... (Data from the previous time step in the multi-source fusion dataset) Output based on the previous time step The urban heat island intensity at the previous moment and the energy balance of the urban canopy together constrain the prediction of the output at the next moment. (Urban heat island intensity at the next moment), the physical rules of urban canopy energy balance play a key constraining role in the prediction of long-term time series data from the previous moment to the next moment. Under the physical rules of urban canopy energy balance, the expression for predicting urban heat island intensity is: This formula is a unique predictive constraint expression created by this invention based on the physical rules of energy balance in urban canopies, for example in... Figure 2 middle, (and The same symbolic representation is... The data corresponding to the first moment represents the temperature change in the urban area at the first moment. (and The same symbolic representation is... The data corresponding to the first moment represents the temperature change in the suburbs surrounding the city at the first moment. Figure 2 The first LSTM unit outputs the heat island intensity. The entire PGRNN fusion model relies on long-term time-series data for training and prediction at each time point. The second LSTM unit predicts the output at the next time point (i.e., the heat island intensity at the second time point) based on the output of the first LSTM unit and the combined constraints of the urban canopy energy balance. Similarly, the PGRNN fusion model is based on the output of the previous time step when training and predicting long-term data. The urban heat island intensity at the previous moment and the energy balance of the urban canopy together constrain the prediction of the output at the next moment. (Urban heat island intensity at the next moment).

[0045] The present invention provides the temperature change values ​​of urban areas over time t-1. Temperature variation over time t-1 in suburban areas surrounding the city All are obtained through inversion using the following formula:

[0046] ,in This represents the change in net radiation. is a surface heat storage net flux change amount, is an anthropogenic heat flux change amount, is a proportional coefficient, is a correction coefficient, is an energy redistribution parameter, , is a Stefan-Boltzmann constant, is a surface temperature. When calculating the air temperature change value of the urban area at time t-1 , the parameters of the urban area at time t-1 are obtained according to the above formula respectively, and the inverse calculation is obtained according to the above formula. Similarly, when calculating the air temperature change value of the urban peripheral suburb at time t-1 , the parameters of the urban peripheral suburb at time t-1 are obtained according to the above formula respectively, and the inverse calculation is obtained according to the above formula.

[0047] The surface heat storage net flux change amount of the present application is calculated and obtained by using multi-source fusion data set according to the following expression through an objective hysteresis model (OHM):

[0048] , wherein , , are urban empirical parameters respectively; is a net radiation amount, is time.

[0049] The anthropogenic heat flux change amount of the present application is an anthropogenic heat flux , which is calculated and obtained by the following formula:

[0050] , wherein is a population density, which is calculated based on a LandScan database, is a total population.

[0051] The urban heat island intensity at time t-1 is expressed as follows:

[0052] , wherein is temperature data of the urban area at time t-1 (the temperature data is selected from air temperature or surface temperature of the urban area at time t-1), is temperature data of the urban peripheral suburb at time t-1 (the temperature data is selected from air temperature or surface temperature of the urban peripheral suburb at time t-1).

[0053] The PGRNN fusion model is verified and constrained by comparing and calculating the root mean square error between the temperature measured value at the weather station and the temperature prediction value output by the PGRNN fusion model at the weather station.

[0054] The joint loss constraint function expression of the PGRNN fusion model is as follows:

[0055] , The mean square error of the output prediction value of the recurrent neural network RNN and the true value, or the mean square error calculated by comparing the temperature measured value at the weather station with the temperature prediction value output by the recurrent neural network RNN in the PGRNN fusion model at the weather station; The error between the output prediction value of the PGRNN fusion model and the true value of the urban heat island intensity, The dynamic adjustment weight coefficient.

[0056] In some embodiments, the present embodiment can utilize the output results of the WRF-UCM fusion model to pre-train the PGRNN fusion model, use the meteorological field driving data (FNL reanalysis data) and land cover data of the WRF-UCM fusion model as the input data of the PGRNN model, and use the air temperature data obtained based on the WRF-UCM fusion model as the output of the PGRNN fusion model for pre-training. The simulation of the WRF-UCM fusion model cannot perfectly reflect the spatial and temporal distribution of the intra-day heat island intensity under the high-temperature heat wave condition, and cannot perfectly match the field observation results, but it provides the basic physical response of the heat island intensity to specific meteorological driving factors and complex urban underlying surface conditions. Therefore, pre-training the PGRNN fusion model using the simulation data of the WRF-UCM fusion model can enable the neural network to simulate and learn the related physical processes. On this basis, the PGRNN fusion model that has been pre-trained can more effectively capture the general physical properties of the input data, reduce the sample size of the training set, that is, better and more effective training results are obtained through less data; this enables the present project to use the limited air temperature monitoring data of automatic stations to obtain the spatial and temporal distribution of intra-day urban heat island intensity; the PGRNN fusion model after pre-training is used for formal training. The FNL reanalysis data and land use data are used as the input layer, and the monitoring data obtained by the automatic weather station are used as the output layer, the PGRNN fusion model is further trained and verified, the automatic weather station observation data is divided, 70% is used as the training set, and 30% is used as the verification set; the training data set and the test data set are input into the PGRNN fusion model, the PGRNN fusion model that has been pre-trained can more effectively capture the general physical properties of the input data, reduce the sample size of the training set, that is, better and more effective training results are obtained through less data. During the training process, the model needs to be fine-tuned in combination with the training monitoring results (which can be decomposed by the loss function or verified by the consistency of physical laws); for example, the adaptive weight method in the loss function can be used to improve the accuracy of model prediction; for the problem of non-convergence of physical residual in the training process, the weight of the physical residual is gradually increased or the adaptive weight is used for optimization; for the problem of overfitting of the model in the training process, data enhancement, adding Dropout / L2 regularization, etc. are used to optimize and solve the problem; for the problem of gradient explosion in the training process, gradient clipping is used to optimize and solve the problem; for the problem of deviation of the prediction result from the physical law, the range of the intermediate data is limited by adding an activation function to the output layer to optimize and solve the problem. After the training of the PGRNN fusion model is completed, the present application further extracts and predicts the high-spatial and temporal resolution urban heat island intensity (Urban Heat Island Intensity, UHII) of the target city area by using the model.The trained PGRNN fusion model is used to infer the reanalyzed meteorological data, land use data and the like as input, and output the temperature prediction value at the corresponding spatio-temporal resolution for the to-be-predicted period and region. Through the physical mechanism introduced in the model, the temperature difference between the city and the suburb can be more accurately described, so as to calculate the heat island intensity.

[0057] The present embodiment uses a multi-source fusion data set and a city heat island intensity matching combination sequence data to learn and train the PGRNN fusion model. In order to verify the accuracy of the method, the temperature prediction value of the model on the automatic weather station (city and suburb) is compared with the measured value, and the root mean square error (RMSE) is calculated to evaluate. And select WRF-UCM simulation results and LSTM-RNN model without physical rule guidance as the baseline to verify.

[0058] The temperature difference between the city site and the corresponding surrounding non-city site is used as the observed heat island intensity. By comparing the UHII value between the model prediction and the observation, the absolute deviation and the correlation coefficient are calculated to test the ability of the model to describe the urban heat island phenomenon. The comparison results show that the PGRNN fusion model of the present application greatly improves the prediction accuracy while retaining the physical consistency, and the spatial details of the heat island intensity are more abundant and the time change response is more timely.

[0059] A city heat island intensity prediction system based on a physical mechanism guided recurrent neural network, comprising a multi-source fusion sample data set, a data acquisition module and a PGRNN fusion model, the multi-source fusion sample data set is a multi-source fusion data set and a city heat island intensity matching combination sequence data or a city heat island intensity sequence data, the data acquisition module is used to collect data including land use cover data and meteorological data in real time and construct a multi-source fusion data set; the PGRNN fusion model is constructed by fusing a recurrent neural network RNN and a long short-term memory network LSTM, and the PGRNN fusion model constructs the following city heat island intensity prediction expression: , wherein , are the city heat island intensities at time t and t-1 respectively, is the city area temperature change value at time t-1, is the city surrounding suburb temperature change value at time t-1; the PGRNN fusion model uses the multi-source fusion sample data set for model prediction training.

[0060] In order to verify the effectiveness and accuracy of the city heat island intensity prediction method and the city heat island intensity prediction system in the city heat island intensity prediction, the present embodiment provides the following two implementation cases:

[0061] Case 1: Taking the main urban area and surrounding areas of Hangzhou City as the research object, a typical high-temperature weather process in a certain month is selected as the verification period, and the high-spatial and temporal resolution heat island intensity prediction results of the model in this period are extracted, and part of the results of the research object are as shown in Figure 4 Figure 4 The spatial distribution of the heat island intensity of the main urban area of Hangzhou City predicted by the PGRNN fusion model of the present application at 14:00 of the research month is shown, and the redder the color scale, the stronger the urban heat island effect. It can be observed that there is a significant temperature difference between the densely built-up area of the main urban area and the surrounding suburbs, and the maximum UHI value reaches 4.46℃, reflecting the good spatial resolution capability of the present application.

[0062] Case 2: The present application selects the automatic weather stations in the urban and suburban areas of Hangzhou City, compares the predicted values and the measured values, and compares the accuracy with the traditional RNN model according to the method of the present application. The RMSE of the PGRNN fusion model of the present application is 2.67℃, the RMSE of the traditional RNN model is 1.23(±0.32)℃, and the RMSE of the PGRNN fusion model of the present application is 1.02(±0.11)℃; the results show that the PGRNN fusion model of the present application greatly improves the accuracy of the heat island intensity prediction while maintaining physical consistency, especially in high-temperature periods and strong convective weather.

[0063] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.​

Claims

1. A method for urban heat island intensity prediction based on physical mechanism guided recurrent neural network, characterized in that: The method comprises the following steps: S1, constructing a multi-source fusion data set comprising land use and cover data and meteorological data; S2, constructing a PGRNN fusion model combining a recurrent neural network (RNN) and a long short-term memory network (LSTM), wherein the PGRNN fusion model is used to predict and train a combination sequence of urban heat island intensity matching data of the multi-source fusion data set and the urban heat island intensity sequence data, and construct the following expression of urban heat island intensity prediction: wherein , are the urban heat island intensity at time t, t-1, respectively, is the urban area temperature change value at time t-1, is the suburban temperature change value at time t-1.

2. The physical mechanism based recurrent neural network guided urban heat island intensity prediction method according to claim 1, characterized in that: Urban area time t-1 air temperature change value , urban area time t-1 air temperature change value are obtained by inversion using the following equation: wherein is the net radiation change, is the net surface heat storage change, is the anthropogenic heat flux change, is a proportionality factor, is a correction factor, is an energy redistribution parameter, , is the Stefan-Boltzmann constant, is the surface temperature. 3.The urban heat island intensity prediction method based on physical mechanism guided recurrent neural network according to claim 1, wherein: The PGRNN fusion model constructs the following energy balance formula of the urban canopy: ; where is the net shortwave radiation, is the net longwave radiation; is the sensible heat flux, is the latent heat flux, is the net ground heat flux.

4. The physical mechanism based recurrent neural network guided urban heat island intensity prediction method according to claim 2, wherein: Surface heat storage net flux change The inversion is calculated by using multi-source fusion data sets through an objective lag model according to the following expression: wherein , , are the urban experience parameters, respectively; is the net radiation, is the time; amount of change in anthropogenic heat flux anthropogenic heat flux is calculated by the following equation: wherein is the population density, is the total population. 5.The urban heat island intensity prediction method based on physical mechanism guided recurrent neural network according to claim 3, wherein: net shortwave radiation The expression is: ; net longwave radiation The expression is: , sensible heat flux The expression is: ; latent heat flux The expression is: ; where is the average surface albedo, is the shortwave radiation reaching the ground, , are the surface emissivity and atmospheric emissivity, respectively, is the Stefan-Boltzmann constant, , are the air temperature and surface temperature, respectively, is the average air density, is the specific heat of air at constant pressure, is the aerodynamic resistance, is the Bowen ratio.

6. The physical mechanism based recurrent neural network guided urban heat island intensity prediction method according to claim 1, wherein: Time t-1 urban heat island intensity The expression is as follows: wherein is temperature data for the urban area at time t-1, is temperature data for the urban periphery at time t-1.

7. The physical mechanism based recurrent neural network guided urban heat island intensity prediction method according to claim 6, characterized in that: The temperature measured at the meteorological station is compared with the temperature predicted by the PGRNN fusion model at the meteorological station, and the root mean square error is calculated, and the PGRNN fusion model is verified and constrained by comparison and root mean square error. 8.The urban heat island intensity prediction method based on physical mechanism guided recurrent neural network according to claim 1, wherein: The joint loss constraint function expression of the PGRNN fusion model is as follows: , is the mean square error of the predicted value and the true value output by the recurrent neural network RNN, or the mean square error calculated by the temperature measured value of the meteorological station at the meteorological station and the temperature predicted value output by the recurrent neural network RNN in the PGRNN fusion model at the meteorological station; is the error between the predicted value and the true value of the urban heat island intensity output by the PGRNN fusion model, is the dynamic adjustment weight coefficient. 9.The urban heat island intensity prediction method based on physical mechanism guided recurrent neural network according to claim 1, wherein: In the method S1, the meteorological data sources include reanalysis meteorological data and meteorological station monitoring data; and the land use and cover data includes a plurality of land cover type data sets classified according to urban areas and surrounding suburban areas. 10.A system for urban heat island intensity prediction based on physical mechanism guided recurrent neural network, characterized in that: The application relates to a city heat island intensity prediction method based on a PGRNN fusion model, which comprises a multi-source fusion sample data set, a data acquisition module and a PGRNN fusion model, the multi-source fusion sample data set is a multi-source fusion data set matched and combined with city heat island intensity sequence data or city heat island intensity sequence data, the data acquisition module is used for collecting data containing land use cover data and meteorological data in real time and constructing a multi-source fusion data set; the PGRNN fusion model is constructed by fusing a recurrent neural network (RNN) and a long short-term memory network (LSTM), and the PGRNN fusion model constructs a city heat island intensity prediction expression as follows: Wherein , are city heat island intensities at times t and t-1 respectively, is a city area temperature change value at time t-1, is a city surrounding suburb temperature change value at time t-1; the PGRNN fusion model is used for model prediction training by using the multi-source fusion sample data set.

Images