Urban population space-time dynamic distribution prediction method based on deep learning

Abstract

The invention discloses an urban population space-time dynamic distribution prediction method based on deep learning, and the method comprises the following steps: 1, predicting a target, dividing the time range of a prediction task in combination with the background of an actual prediction task, dividing the prediction task after n scores, n hours, n days or n weeks in the future into short-term prediction, dividing the prediction task after n months or n years in the future into long-term prediction; 2, data preprocessing; the method comprises the following steps: firstly collecting and sorting multi-source data, and then performing rasterization processing in combination with a geographic information GIS technology; 3, selecting a corresponding prediction model, inputting a result of the previous step, and performing model prediction; and 4, explaining and analyzing a model prediction result by using an SHAP value analysis method to obtain contribution degrees of different features to a population distribution prediction result. On the premise of improving the accuracy of urban population spatial-temporal dynamic distribution prediction, the problem of spatial dependency of population prediction is solved.

Description

Technical Field

[0001] The present invention relates to the field of traffic management technology, and in particular to a method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning. Background Art

[0002] As the global urbanization process accelerates, urban planning and efficient management of transportation systems are crucial to the sustainable development of cities. The spatiotemporal dynamic distribution of urban population is the core element of urban planning and transportation demand analysis, which is related to the rational allocation of resources and the quality of life and environmental sustainability of cities.

[0003] Existing urban population forecasting methods face multiple challenges, including the need for fine-grained spatiotemporal forecasting, spatial dependency issues, and model transferability and generalization issues.

[0004] Gridded population data from census sources are typically updated annually, which makes it challenging to infer population dynamics within a year and difficult to conduct fine-grained spatiotemporal population forecasts.

[0005] Secondly, it is difficult for traditional population forecasting methods to take into account the spatial heterogeneity of population distribution within administrative units while refining it to the grid level.

[0006] However, existing grid allocation methods based on auxiliary geographic datasets such as night lights and land cover can hardly reflect the relationship between urban built environment and population, and ignore the impact of different types of areas and transportation accessibility on population agglomeration.

[0007] In this context, how to accurately predict the spatiotemporal dynamic distribution data of urban population has become a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0008] In view of the above-mentioned defects of the prior art, the present invention provides a method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning. The purpose of the present invention is to improve the accuracy of the prediction of the spatiotemporal dynamic distribution of urban population, further meet the needs of fine-grained spatiotemporal estimation, solve the problem of spatial dependence of population prediction, and provide a transferable and generalizable prediction method.

[0009] To achieve the above object, the present invention discloses a method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning, comprising the following steps:

[0010] Step 1: Predict the target, specifically:

[0011] In combination with the actual background of the forecasting task, the time range of the forecasting task is divided, and the forecasting task after n minutes, n hours, n days or n weeks in the future is divided into short-term forecasting, and the forecasting task after n months or n years in the future is divided into long-term forecasting; n is a natural number greater than 1;

[0012] Step 2: Data preprocessing, specifically:

[0013] First, we collect and organize multi-source data including historical population data, weather, holidays, emergencies, infrastructure POIs and road networks, and then use geographic information GIS technology to perform raster processing as input data for the next prediction model;

[0014] Step 3: Select the corresponding prediction model, input the result of the previous step, and perform model prediction;

[0015] Step 4: Use SHAP, i.e., SHapley Additive explanation value analysis method, to interpret and analyze the results predicted by the model to obtain the contribution of different features to the population distribution prediction results.

[0016] Preferably, in step 2, the historical population data includes but is not limited to population data obtained based on mobile phone signaling data and individual trajectory data.

[0017] More preferably, in step 2, for the forecasting task classified as the short-term forecasting, external information data including the historical population data and weather, holidays, emergencies, and infrastructure are first prepared;

[0018] The granularity of the original basic data of the external information data is aligned with the granularity of the prediction task;

[0019] Then, the original basic data of the external information data is rasterized in combination with the geographic information GIS technology to obtain raster data of d dimensions; d is the number of data types;

[0020] Select the long-term, medium-term and short-term d-dimensional grid data combinations from the historical population data as feature data, which are used as input data for the next prediction model;

[0021] Among them, the same moment one week ago is the long term; the same moment one day ago is the medium term; t moments ago is the recent term, and the feature data is the grid data of (d+2+t) dimension.

[0022] More preferably, in step 3, when predicting the spatiotemporal dynamic distribution data of the population in a single region, the EFLSTM model is used to perform the model prediction;

[0023] When predicting multiple regions and considering the spatiotemporal dynamic distribution data of the population of regional interactions, the EFCNN-LSTM model is used for the model prediction.

[0024] More preferably, in step 2, for the prediction task classified as the long-term prediction, the historical population data, infrastructure POI data and road network data are first prepared, and then the road network data is used to perform accessibility analysis to obtain the accessibility area ranges of 10 minutes, 20 minutes, 30 minutes, 40 minutes, 50 minutes, 60 minutes, 75 minutes and 90 minutes;

[0025] The infrastructure POI data includes government department data, workplace data, shopping data, education data, tourist attraction data, fitness data, residence data, hospital data and other data;

[0026] Then, the infrastructure POI data is integrated with the result of accessibility analysis using the road network data to obtain POI attributes within reachable ranges at different times;

[0027] The POI attributes within the reachable range at different times include 9 types of POIs×9 types of arrival times=81 attribute values;

[0028] Combined with the geographic information GIS technology, rasterization processing is performed to obtain 81-dimensional feature data as input data for the next prediction model.

[0029] More preferably, in step 3, a random forest prediction model is used, and the feature data of 81 dimensions is input and then the model prediction is performed; wherein 80% of the data is a training set and 20% of the data is a test set, so as to obtain a long-term population spatiotemporal dynamic distribution prediction result.

[0030] Preferably, in step 4, the SHAP value analysis method is used to perform an explanatory analysis on the results of the model prediction, to explore the influence of different input feature data on the model prediction, to reveal the contribution of each feature dimension, and to increase the interpretability of the prediction model.

[0031] Beneficial effects of the present invention:

[0032] On the premise of improving the accuracy of prediction of spatiotemporal dynamic distribution of urban population, the present invention further meets the need of fine-grained spatiotemporal estimation, solves the problem of spatial dependence of population prediction, and provides a transferable and generalizable prediction method.

[0033] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The following is a flowchart showing an implementation of an embodiment of the present invention. DETAILED DESCRIPTION

[0035] Example

[0036] like Figure 1 As shown in FIG, the method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning includes the following steps:

[0037] Step 1: Predict the target, specifically:

[0038] According to the actual prediction task background, the time range of the prediction task is divided into short-term prediction tasks after n minutes, n hours, n days or n weeks in the future, and long-term prediction tasks after n months or n years in the future; n is a natural number greater than 1;

[0039] Step 2: Data preprocessing, specifically:

[0040] First, we collect and organize multi-source data including historical population data, weather, holidays, emergencies, infrastructure POIs and road networks, and then use geographic information GIS technology to perform raster processing as input data for the next prediction model;

[0041] Step 3: Select the corresponding prediction model, input the result of the previous step, and perform model prediction;

[0042] Step 4: Use SHAP, the SHapley Additive explanation value analysis method, to interpret and analyze the results of the model prediction and obtain the contribution of different features to the population distribution prediction results.

[0043] The present invention comprehensively utilizes historical population data, point of interest (POI) data, road network topology data, external weather, whether it is a holiday and other multi-source data, and based on deep learning algorithm technology, develops a population spatiotemporal dynamic distribution prediction model with fine-grainedness and high spatiotemporal resolution, and proposes a method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning to improve the accuracy of population distribution prediction data.

[0044] In some embodiments, in step 2, the historical population data includes, but is not limited to, population data obtained based on mobile phone signaling data and individual trajectory data.

[0045] In some embodiments, in step 2, for the prediction task classified as short-term prediction, external information data including historical population data and weather, holidays, emergencies, and infrastructure are first prepared;

[0046] The granularity of the original basic data of the external information data is aligned with the granularity of the prediction task;

[0047] For example, if the task goal is to predict the spatiotemporal distribution of the population after n minutes, the granularity of the original basic data of the external information data that needs to be prepared should be at least at the n-minute level.

[0048] The original basic data of the external information data is then rasterized using geographic information GIS technology to obtain raster data of d dimensions; d is the number of data types; for example, if the original basic data includes historical population data, weather and holiday data, then the value of d is 3.

[0049] Select the long-term, medium-term and short-term d-dimensional grid data combination from the historical population data as feature data, which will be used as the input data for the next prediction model;

[0050] Among them, the same time a week ago is long-term; the same time a day ago is medium-term; t moments ago is short-term, so the feature data is raster data of (d+2+t) dimension.

[0051] In some embodiments, in step 3, when predicting the spatiotemporal dynamic distribution data of the population in a single region, the EFLSTM model is used for model prediction;

[0052] When predicting multiple regions and considering the spatiotemporal dynamic distribution data of population in relation to regional interactions, the EFCNN-LSTM model is used for model prediction.

[0053] In practical applications, when predicting the spatiotemporal dynamic distribution data of the population in a single region, the EFLSTM model is selected for the prediction task with higher accuracy; when predicting the spatiotemporal dynamic distribution data of the population in multiple regions (considering regional interactions), the EFCNN-LSTM model is selected for the prediction task with higher accuracy. The raster data processed in the second step is input into the model for prediction (80% of the data is the training set and 20% of the data is the test set) to obtain the prediction results of the short-term spatiotemporal dynamic distribution of the population.

[0054] In some embodiments, in step 2, for the prediction task classified as long-term prediction, historical population data, infrastructure POI data and road network data are first prepared, and then accessibility analysis is performed using the road network data to obtain accessibility area ranges of 10 minutes, 20 minutes, 30 minutes, 40 minutes, 50 minutes, 60 minutes, 75 minutes and 90 minutes;

[0055] Infrastructure POI data includes government department data, workplace data, shopping data, education data, tourist attraction data, fitness data, residence data, hospital data and other data.

[0056] Then, the infrastructure POI data is integrated with the results of accessibility analysis using road network data to obtain the POI attributes within the reachable range at different times;

[0057] The attributes of POIs within reachable ranges at different times include 9 types of POIs × 9 types of arrival times = 81 attribute values;

[0058] Combined with geographic information GIS technology, rasterization processing is performed to obtain 81-dimensional feature data, which serves as the input data for the next prediction model.

[0059] In some embodiments, in step 3, a random forest prediction model is used, and 81-dimensional feature data is input and then model prediction is performed; among them, 80% of the data is a training set and 20% of the data is a test set, and the long-term population spatiotemporal dynamic distribution prediction results are obtained.

[0060] In some embodiments, the SHAP value analysis method is used to perform an explanatory analysis on the results of model prediction, explore the influence of different input feature data on model prediction, reveal the contribution of each feature dimension, and increase the interpretability of the prediction model.

[0061] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning; characterized in that: The steps include: Step 1: Predict the target, specifically: In combination with the actual background of the forecasting task, the time range of the forecasting task is divided, and the forecasting task after n minutes, n hours, n days or n weeks in the future is divided into short-term forecasting, and the forecasting task after n months or n years in the future is divided into long-term forecasting; n is a natural number greater than 1; Step 2: Data preprocessing, specifically: First, we collect and organize multi-source data including historical population data, weather, holidays, emergencies, infrastructure POIs and road networks, and then use geographic information GIS technology to perform raster processing as input data for the next prediction model; Step 3: Select the corresponding prediction model, input the result of the previous step, and perform model prediction; Step 4: Use SHAP, i.e., SHapley Additive explanation value analysis method, to interpret and analyze the results predicted by the model to obtain the contribution of different features to the population distribution prediction results.

2. The method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning according to claim 1 is characterized in that: In step 2, the historical population data includes but is not limited to population data obtained based on mobile phone signaling data and individual trajectory data.

3. The method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning according to claim 2 is characterized in that: In step 2, for the forecasting task classified as the short-term forecasting, external information data including the historical population data and weather, holidays, emergencies, and infrastructure are first prepared; The granularity of the original basic data of the external information data is aligned with the granularity of the prediction task; Then, the original basic data of the external information data is rasterized in combination with the geographic information GIS technology to obtain raster data of d dimensions; d is the number of data types; Select the long-term, medium-term and short-term d-dimensional grid data combinations from the historical population data as feature data, which are used as input data for the next prediction model; Among them, the same moment one week ago is the long term; the same moment one day ago is the medium term; t moments ago is the recent term, and the feature data is the grid data of (d+2+t) dimension.

4. The method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning according to claim 3 is characterized in that: In step 3, when predicting the spatiotemporal dynamic distribution data of the population in a single region, the EFLSTM model is used to perform the model prediction; When predicting multiple regions and considering the spatiotemporal dynamic distribution data of the population of regional interactions, the EFCNN-LSTM model is used for the model prediction.

5. The method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning according to claim 2 is characterized in that: In step 2, for the prediction task classified as the long-term prediction, the historical population data, infrastructure POI data and road network data are first prepared, and then the road network data is used to perform accessibility analysis to obtain the accessibility area ranges of 10 minutes, 20 minutes, 30 minutes, 40 minutes, 50 minutes, 60 minutes, 75 minutes and 90 minutes; The infrastructure POI data includes government department data, workplace data, shopping data, education data, tourist attraction data, fitness data, residence data, hospital data and other data; Then, the infrastructure POI data is integrated with the result of accessibility analysis using the road network data to obtain POI attributes within reachable ranges at different times; The POI attributes within the reachable range at different times include 9 types of POIs×9 types of arrival times=81 attribute values; Combined with the geographic information GIS technology, rasterization processing is performed to obtain 81-dimensional feature data as input data for the next prediction model.

6. The method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning according to claim 5 is characterized in that: In step 3, the Random Forest prediction model is used to input the 81-dimensional feature data and perform the model prediction; 80% of the data is a training set and 20% of the data is a test set, and the long-term population spatiotemporal dynamic distribution prediction result is obtained.

7. The method for predicting the spatiotemporal dynamic distribution of urban population based on deep learning according to claim 1 is characterized in that: In step 4, the SHAP value analysis method is used to perform an explanatory analysis on the results predicted by the model, explore the influence of different input feature data on the model prediction, reveal the contribution of each feature dimension, and increase the interpretability of the prediction model.

Images