Urban employment scale prediction method based on dynamic simulation and feedback calibration

Abstract

An urban employment scale prediction method based on dynamic simulation and feedback calibration is characterized by comprising the following steps: S1, data acquisition and processing: acquiring productivity data and employment data of a target industry in a plurality of historical census years, and calculating a historical comprehensive growth factor sequence based on the data; s2, constructing and calibrating a dynamic simulation twinborn model; s3, performing initial prediction; s4, executing twinning trend simulation; s5, trend comparison and deviation degree analysis: generating benchmark employment trend data based on the historical employment data, comparing the simulated employment trend data obtained in the step S4 with the benchmark employment trend data, and calculating a trend deviation degree between the simulated employment trend data and the benchmark employment trend data; and S6, performing feedback calibration and output. The method has the beneficial effects that the prediction robustness and the self-adaptive capability are remarkably improved, and the mechanism property and the interpretability are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, computer simulation and intelligent prediction, and particularly relates to a city employment scale prediction method based on dynamic simulation and feedback calibration. BACKGROUND

[0002] City employment scale prediction is an important basis for city planning and policy making. Existing prediction methods mainly include three categories:

[0003] Classical statistical regression methods: such as linear regression, establish the relationship between economic, population and other independent variables and employment dependent variables. The disadvantage is that the relationship is assumed to be linear and static, which is difficult to describe complex nonlinear interaction and dynamic change.

[0004] Time series analysis methods: such as ARIMA model, based on historical data trend extrapolation. The disadvantage is that the response to sudden external shocks (such as major policy adjustments, economic crisis) is lagging, and the prediction is prone to large deviations.

[0005] Machine learning methods: such as neural networks, random forests, which can capture nonlinear patterns. The disadvantage is that the model is less interpretable, and most of them are "black box" predictions, lacking the mechanism of "why such prediction". When the external environment changes dramatically, the internal logic of the model cannot be adjusted automatically, and the generalization ability is limited.

[0006] The closest prior art: Chinese patent with authorization announcement number CN118504775A discloses a city planning method based on digital twinning, which detects population data anomalies by improving the isolated forest algorithm, and then optimizes the prediction. This technology improves the accuracy of abnormal data processing, but the role of digital twinning is mainly reflected in the data cleaning and preprocessing stage, not in the core modeling and dynamic calibration of the prediction process. Its prediction model itself (such as exponential smoothing method) is still static, with fixed parameters, and cannot be self-corrected according to the new trends found in the simulation during the prediction process.

[0007] Therefore, in the existing technology, especially for city planning prediction models, there is a common technical defect: model parameters are fixed, and there is a lack of a closed-loop mechanism that can simulate the complex dynamics of the real world and use the simulation results for real-time calibration of the model itself. This leads to a serious decline in the stability and accuracy of the prediction results when facing non-stationary and nonlinear effects caused by industrial structure adjustment, major policy implementation, etc. SUMMARY

[0008] The present application aims to solve the technical problems of poor prediction robustness and insufficient accuracy in existing city employment scale prediction methods due to the static model that cannot perceive and respond to external dynamic complex factors.

[0009] In order to achieve the above-mentioned purpose of the application, the application provides the following technical solutions:

[0010] In order to achieve the above-mentioned purpose, the application adopts the following principles: a city employment scale dynamic prediction method based on digital twin simulation feedback, the core of which is to build a closed loop system of "prediction-simulation-comparison-calibration".

[0011] Firstly, an initial prediction model (such as a regression model combining capacity factor and growth factor) is established for preliminary prediction.

[0012] The key innovation is to input the preliminary prediction result into a pre-constructed dynamic simulation twin model (such as a system dynamics model) that can reflect the internal mechanism of industry operation, and obtain detailed employment evolution trend in the future period through simulation.

[0013] Then, the simulation trend is intelligently compared with the benchmark trend obtained based on pure historical data statistics, and the "trend deviation degree" is quantified.

[0014] Finally, a feedback calibration mechanism is introduced: when the deviation degree exceeds the threshold, the system automatically determines that there is a significant inconsistency between the initial prediction model and the real world dynamics, and then triggers an optimization algorithm to adjust the parameters (such as the calculation formula coefficients of the growth factor) of the initial model, so that the new prediction made by the adjusted model can be simulated by the twin model, and its output trend can be maximized. The benchmark trend is consistent. Through this closed loop, the prediction model has the ability to "learn" and "adapt".

[0015] The specific technical solutions are as follows:

[0016] A city employment scale prediction method based on dynamic simulation and feedback calibration, characterized by the following steps:

[0017] S1, obtaining the capacity data and employment data of the target industry in multiple historical census years, and calculating the historical comprehensive growth factor sequence based on the data;

[0018] S2, constructing and calibrating a dynamic simulation twin model: using system dynamics method, establishing a dynamic simulation twin model with capacity stock and employment stock as state variables; based on the historical capacity, employment data and annual investment sequence estimated from the historical data, determining an objective search space for model parameters, and performing parameter calibration through optimization algorithm to obtain a verified model that can reproduce historical dynamics;

[0019] S3, initial prediction: predicting the industry capacity and comprehensive growth factor of the target year respectively, and calculating the first employment prediction value based on the preset employment prediction equation;

[0020] S4. Perform twin trend simulation: Use the first employment forecast value as a constraint on future employment, input it into the verified dynamic simulation twin model, and run the simulation to obtain simulated employment trend data for the target year and the subsequent preset period.

[0021] S5. Trend Comparison and Deviation Analysis: Generate benchmark employment trend data based on the historical employment data, compare the simulated employment trend data obtained in step S4 with the benchmark employment trend data, and calculate the trend deviation between the two.

[0022] S6. Feedback calibration and output: Determine whether the trend deviation exceeds a preset threshold; if so, with the goal of reducing the deviation, reverse-calibrate the parameters used to predict the comprehensive growth factor in step S3, recalculate the comprehensive growth factor and employment scale using the calibrated parameters, and output the final employment prediction value; if not, directly output the first employment prediction value as the final prediction result.

[0023] Furthermore, in step S2, the core dynamic equations of the dynamic simulation twin model are as follows:

[0024]

[0025]

[0026] in, and They represent the first Annual production capacity and employment stock; and They represent the first Annual fixed asset investment and comprehensive growth factors; The model parameters to be calibrated represent the investment conversion efficiency coefficient, the capacity-to-employment growth coefficient, the influence intensity of the growth factor, the capacity depreciation rate, and the natural job loss rate, respectively.

[0027] Furthermore, in step S2, determining the objective search space for the model parameters specifically includes:

[0028] The capacity employment-driving coefficient The search scope is determined by calculating the instantaneous ratio of employment to production capacity for each historical census year. And based on the minimum and maximum values ​​of this ratio sequence, it is extended to determine;

[0029] The investment conversion efficiency coefficient The search scope is determined by calculating efficiency estimates for multiple periods based on the increase in production capacity and the estimated total investment during the same period between adjacent historical census years, and then expanding the scope based on the minimum and maximum values ​​of the estimated value sequence.

[0030] the capacity depreciation rate and the natural attrition rate of employment The search range of the parameters is determined according to the reasonable range published by the relevant asset depreciation regulations of the country and the industry labor mobility statistical report, respectively.

[0031] Further, in step S2, the parameter calibration by the optimization algorithm is specifically: a particle swarm optimization algorithm is used to minimize the weighted root mean square error between the census year state value of the model simulation output and the corresponding historical observation value as the fitness function, and iterative optimization is performed in the determined parameter search space to obtain the optimal parameter set.

[0032] Further, in step S3, the first employment prediction value is calculated based on the preset employment prediction equation, and the equation is:

[0033]

[0034] wherein, is the predicted number of employment in the target year n, is the predicted capacity in the target year n, is the predicted comprehensive growth factor in the target year n; are the number of employment and the capacity of the historical census year immediately before the target year, respectively.

[0035] Further, in step S3, the prediction of the industry capacity and the comprehensive growth factor in the target year is achieved by performing nonlinear regression fitting on the corresponding historical data sequence, respectively.

[0036] Further, in step S6, the parameters for predicting the comprehensive growth factor are calibrated in reverse, specifically: the form of the prediction model is kept unchanged, the minimization of the trend deviation degree is taken as the target, and the one-dimensional search or gradient descent method is used to adjust the parameters of the prediction model until the new deviation degree of the simulated trend data generated after the simulation of the employment value calculated by the growth factor predicted by the new parameters in step S4 and the benchmark trend data is lower than the preset threshold.

[0037] Further, in step S5, the benchmark employment trend data is obtained by performing time series decomposition on the historical employment data sequence and extracting the long-term trend item, or by fitting and extrapolating using the exponential smoothing method; and the trend deviation degree is obtained by calculating the average absolute percentage error or the root mean square error of the simulated employment trend data and the benchmark employment trend data in the same time interval.

[0038] Further, the method is cyclically executed in batch mode, and a plurality of different industries are independently predicted in turn, wherein the model construction and rate setting of step S2 are independently completed for each industry using its own historical data.

[0039] The application also includes an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of the preceding method embodiments when executing the program.

[0040] The application also includes a computer readable storage medium having a computer program stored thereon, wherein the program implements the method according to any one of the preceding method embodiments when executed by a processor.

[0041] Compared with the prior art, the application has the following beneficial technical effects:

[0042] The robustness and adaptability of the prediction are significantly improved: through the feedback closed loop of the twin simulation, the prediction model can dynamically perceive external disturbances such as policies and economies, and automatically adjust its parameters to adapt to changes, solving the problem of failure of static models in non-stationary scenarios.

[0043] The mechanism and interpretability of the prediction process are enhanced: the dynamic simulation twin model is based on mechanism modeling methods such as system dynamics, which can simulate the internal logic of how changes in production capacity are transmitted to employment, so that the prediction is no longer a "black box" result driven by data alone, but is supported by process simulation.

[0044] The prediction accuracy is substantially improved: by feeding back and calibrating the model in advance with the potential future trend risks (such as weak growth or sharp fluctuations) discovered by simulation, systematic biases generated by simply relying on historical data extrapolation can be effectively corrected, thereby obtaining more accurate final prediction values. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The total flowchart of the closed-loop dynamic prediction method provided by the embodiment of the application.

[0046] Figure 2 The principle diagram of the trend comparison and feedback calibration process in the embodiment of the application. DETAILED DESCRIPTION

[0047] As Figure 1As shown, it is a complete flowchart of the urban employment scale prediction method based on dynamic simulation and feedback calibration provided by the embodiment of the present application. As shown in the figure, the method of the present application forms a logical closed loop containing six core steps: S1 data preparation → S2 model building and rating → S3 initial prediction → S4 twin trend simulation → S5 trend comparison analysis → S6 feedback calibration and output. It is particularly worth noting that the output result (calibration instruction) of step S6 will return and affect the prediction parameters of step S3, thereby forming a self-optimizing and continuously improving “feedback calibration loop” (highlighted in red dashed arrow in the figure). The flowchart clearly structures the overall technical context of the present application, which will be described in detail below in conjunction with each step.

[0048] The present embodiment takes the prediction of the employment scale of the manufacturing industry in Beijing in 2028 as an example to demonstrate the complete technical process. All original data strictly uses economic census data.

[0049] Table 1: Data of production capacity of each industry in Beijing in 2004-2023 from each economic census

[0050] Industry Output K (billion yuan) Year 2004 2008 2013 2018 2023 Total (excluding agriculture, forestry, animal husbandry and fishery) 216316.88 646672.97 1220785.22 2119533.43 2688117.80 Mining industry 68.95289 433.12 2927.11 3983.70 3957.70 Manufacturing industry 6531.14764 9981.83 17827.37 26713.70 37259.30 Electricity, heat, gas and water production and supply industry 6851.7584 7317.27 13204.38 21867.40 34998.40 Construction industry 3199.9277 7283.40 18021.11 35232.60 39441.10 Wholesale and retail industry 8102.98416 17832.89 40399.02 62339.80 93574.70 Transport, warehouse and postal industry 2494.31322 4999.95 15176.94 54752.43 59911.80 Accommodation and catering industry 777.65868 1204.57 2048.57 2559.20 2998.80 Information transmission, software and information technology service industry 8734.77205 11120.28 28206.22 55834.30 96209.30 Financial industry 141269.933 474100.41 873336.48 1486292.50 1827687.60 Real estate industry 10016.35881 20042.31 47129.88 96972.50 121405.60 Leasing and business service industry 19440.64888 74670.77 117957.28 194326.70 267532.80 Scientific research and technology service industry 2932.32279 9569.24 24304.59 43862.10 58241.40 Water conservancy, environment and public facilities management industry 713.98917 929.30 2870.89 5839.40 8736.70 Residents' services, repair and other service industry 822.70847 252.69 897.70 1075.30 1100.40 Education 1120.79147 1990.27 3710.89 6591.70 6868.00 Health and social work 464.5819 741.88 1338.16 2513.10 4139.70 Culture, sports and entertainment industry 1083.17422 1746.99 4367.28 9011.10 11210.00 Public management, social security and social organization 1690.85531 2455.78 7061.37 9765.90 12844.50

[0051] Table 2: Data of employment number of each industry in Beijing in 2004-2023 from each economic census

[0052] Industry Employment E (million people) Year 2004 2008 2013 2018 2023 Total (excluding agriculture, forestry, animal husbandry and fishery) 705.15 816.86 1111.34 1360.85 1313.30 Mining industry 4.14 5.14 6.87 3.57 2.3 Manufacturing industry 148.79 134.69 138.55 96.84 81.2 Electricity, heat, gas and water production and supply industry 7.79 6.73 9.26 10.78 10.4 Construction industry 69.68 54.70 65.52 96.45 92.1 Wholesale and retail industry 83.22 94.34 147.77 161.40 136 Transport, warehouse and postal industry 43.00 69.54 68.57 71.46 56.8 Accommodation and catering industry 38.37 44.95 50.59 56.27 58.2 Information transmission, software and information technology service industry 28.51 46.64 93.00 138.88 160.6 Financial industry 15.00 25.13 43.28 80.56 74.5 Real estate industry 31.86 40.99 56.35 73.02 68.8 Leasing and business service industry 63.78 95.85 141.73 187.22 188.6 Scientific research and technology service industry 38.87 56.44 95.24 140.43 133.4 Water conservancy, environment and public facilities management industry 7.45 9.28 12.87 18.39 18.1 Residents' services, repair and other service industry 19.57 14.10 21.47 30.60 33.7 Education 38.71 43.08 53.07 67.10 64 Health and social work 16.05 19.97 27.81 36.35 43.2 Culture, sports and entertainment industry 16.22 18.64 27.60 37.46 39.2 Public management, social security and social organization 34.13 36.65 51.77 54.07 52.2

[0053] Step S1: Data acquisition and processing

[0054] The historical data of the manufacturing industry is extracted from Table 1 and Table 2, as shown in Table 3:

[0055] Economic census year (t) 2004 2008 2013 2018 2023 Capacity (Billion Yuan) 6531.15 9981.83 17827.37 26713.70 37259.30 Employment (per 1,000 people) 148.79 134.69 138.55 96.84 81.20

[0056] First, according to the relationship between production capacity and employment, the historical growth factor between adjacent census years is calculated . This factor comprehensively reflects the impact of all other factors such as technology and policy on employment in addition to the direct pull of production capacity.

[0057]

[0058] Among them, and represent the employment number of year t+1, year t, respectively, and represent the production capacity of year t+1, year t, respectively. Thus, the growth factor of 2008, 2013, 2018, 2023 relative to the previous census year can be deduced:

[0059]

[0060] Similarly, we have:

[0061]

[0062] So far, we have obtained the core data series needed for modeling and forecasting: .

[0063] Secondly, to build a dynamic simulation model, we need to estimate the annual fixed asset investment series Based on the classic economic assumption that "new production capacity mainly comes from investment", we can use a simplified formula to estimate it. For example, estimate the investment in 2005

[0064]

[0065] Where the constant 0.5 is the assumed average capital output ratio. Through this method, we can linearly interpolate the investment estimation series for each year from 2004 to 2023 .

[0066] Step S2: Construction and parameter calibration of dynamic simulation twin model

[0067] 2.1 Model structure definition

[0068] Build a system dynamics model containing two core state variables: production capacity stock and employment stock.

[0069] State variables:

[0070] : Production capacity stock in year t (billion yuan).

[0071] : Employment stock in year t (ten thousand people). Core dynamic equations:

[0072]

[0073]

[0074] Where:

[0075] : Investment in year t.

[0076] : Growth factor in year t, which is interpolated from its historical series in non-census years.

[0077] : Employment in year t (ten thousand people).

[0078] ​​​​ are the model core parameters to be calibrated, whose physical meanings are:

[0079] : investment conversion efficiency coefficient, representing the effective production capacity formed per unit of investment.

[0080] : production capacity employment pull coefficient, representing the direct employment that can be pulled by per unit of production capacity stock.

[0081] : growth factor influence intensity, representing the amplification or inhibition degree of the comprehensive growth factor on employment changes.

[0082] : production capacity depreciation rate, representing the annual natural loss rate of production capacity due to factors such as technology elimination and equipment aging.

[0083] : employment natural loss rate, representing the annual natural reduction rate of employment due to factors such as retirement and career change.

[0084] 2.2 Objective determination of parameter search space

[0085] To ensure the efficiency and reliability of the calibration results, the search ranges of all parameters are objectively determined based on public historical data and statistical laws, rather than arbitrarily set.

[0086] The range of production capacity employment pull coefficient β: This parameter is directly related to the static proportion of production capacity and employment. Calculate the instantaneous ratio for each census year:

[0087]

[0088] Substitute the manufacturing data to calculate the sequence: {0.0228, 0.0135, 0.0078, 0.0036, 0.0022}. This sequence shows a downward trend, reflecting the historical law of "reduction of labor required per unit of production capacity" brought about by technological progress. Based on the minimum value 0.0022 and the maximum value 0.0228, expand 20% to both ends to cover potential fluctuations, and get the search range:

[0089]

[0090] The range of investment conversion efficiency coefficient α: This parameter is related to the ratio of new production capacity and investment. Based on the simplified model "new production capacity ≈ effective investment", use the data of adjacent census years to estimate:

[0091]

[0092] where Δ is the interval between censuses (4 or 5), The annual investment estimated in step S1. Calculate the α estimate value of multiple interval periods. Take the minimum and maximum of these estimates and expand 15% as the search range, for example:

[0093]

[0094] Capacity depreciation rate and the range of the employment loss rate :

[0095] Mainly reflects the depreciation of fixed assets. According to the provisions of the "Enterprise Income Tax Law of the People's Republic of China" on the minimum depreciation period of machinery and equipment (10 years) and the depreciation period of buildings (20 years), the corresponding annual depreciation rate theoretical range is 5% to 10%. Combined with the actual situation of the manufacturing industry, set a conservative range:

[0096]

[0097] Reflects the natural liquidity of the labor market. Referring to the "Beijing Manufacturing Employment Flow Report" published by the Beijing Statistics Bureau, the "resignation rate" data is usually between 8% and 15%. According to this, set the range:

[0098]

[0099] The range of the influence strength γ of the growth factor: this parameter adjusts the impact strength of external comprehensive factors on employment. The historical growth factor The absolute value of the sequence {-0.335, 0.107, -0.467, -0.257} is less than 0.5. In order to make the model simulate historical fluctuations but avoid over-response, set γ to fluctuate around 1, that is:

[0100]

[0101] 2.3 Parameter calibration process

[0102] The goal of calibration is to find a set of optimal parameters , so that the model starts from the initial state in 2004 ( ), driven by and , runs to 2023, and the overall error of the census year state value of the output and the historical observation value , is the smallest.

[0103] Initialization: Use Latin hypercube sampling method to generate 100 sets of initial parameter combinations in the search space of each parameter determined above to ensure that the search space is evenly covered.

[0104] Iterative optimization: Particle Swarm Optimization (PSO) algorithm is employed for automatic calibration.

[0105] Fitness function: defined as the weighted Root Mean Square Error (RMSE) of the model simulation output and historical data across all census years, to match the dynamics of both capacity and employment simultaneously:

[0106]

[0107] where, The weight coefficient 0.5 indicates equal importance of fitting accuracy for capacity and employment.

[0108] Algorithm execution: set the PSO population size to 50 and the maximum number of iterations to 200. In each iteration, each particle (i.e. a set of parameters updates its parameters according to its individual historical best position and the global best position of the population, and recalculates the fitness value. The optimization process continues until the fitness value converges or the maximum number of iterations is reached.

[0109] Calibration results and validation:

[0110] After optimization, a set of optimal parameters is obtained:

[0111]

[0112] Model validation: the optimal parameters are substituted into the model, running from 2004 to 2023. Calculate the determination coefficient R^2 of the simulation sequence in the census year and the historical data:

[0113]

[0114]

[0115] The values are all greater than 0.9, indicating that the calibrated model can highly reproduce the core dynamics of the historical development of the manufacturing industry, and the internal mechanism (especially the natural attrition effect of employment) is reliable. The validated model is used as the core engine for trend simulation in the subsequent steps.

[0116] Step S3: Initial prediction and trend simulation

[0117] Initial prediction:

[0118] Perform a quadratic polynomial fitting on the sequences and respectively, to obtain the fitting function ​and .

[0119] Predict the value in 2028 (t=2028):

[0120]

[0121] Plug in the initial prediction equation to calculate the first employment prediction value:

[0122]

[0123] Trend simulation:

[0124] As the target endpoint for employment in 2028, combine the capacity prediction path , and assume that investments are generated in a certain proportion of the capacity growth rate from 2024 to 2028. Input the above predicted future input sequence into the dynamic simulation twin model calibrated in step S2, and run from 2023 as the starting point

[0125] to 2028.

[0126] The model calculates according to its internal dynamic equation (especially = 0.102 loss rate) and outputs a simulated employment trend curve from 2024 to 2028 . The simulation results show that due to the strong natural loss effect, the employment growth is inhibited, and the simulation value = 821,000 people in 2028, lower than the initial prediction of 875,000 people. This step corresponds to the arrow path from step S3 to step S4 in the Figure 1 .

[0127] Step S4: Trend Comparison and Deviation Calculation

[0128] Generate baseline trend: apply exponential smoothing method (set smoothing coefficient to 0.3) to extrapolate the baseline trend curve from 2024 to 2028 .

[0129] Calculate trend deviation: calculate the average absolute percentage error of the two curves and in the period from 2024 to 2028 as the quantitative trend deviation D. The left half of the Figure 2 (trend comparison chart) intuitively shows the comparison relationship between the simulated employment trend curve (red triangular dotted line) and the baseline trend curve (green dotted line) in the prediction interval (gray area), as well as the trend deviation D calculated therefrom.​​

[0130]

[0131] Assume the calculation of .

[0132] Step S5: Feedback calibration and final prediction

[0133] Threshold judgment: preset calibration threshold . Because , the system determines that the initial prediction does not match the simulation trend, triggering feedback calibration. The above judgment logic is Figure 1 the process from step S3 to step S4 in the above.

[0134] Parameter reverse optimization:

[0135] Establish an optimization problem: take the minimum deviation D as the goal, and take the coefficient of the growth factor fitting function in the initial prediction as the parameter to be adjusted.

[0136] Use one-dimensional search method for adjustment. For example, keep a, b unchanged, and only adjust the constant term c to directly change .

[0137] For each candidate :

[0138] a. Calculate the new employment prediction value .

[0139] b. Take as the new target in 2028, call the twin model again for simulation, and get the new curve .

[0140] c. Calculate the new deviation D'.

[0141] After optimization, the calibrated growth factor that makes D' minimum is obtained. The right half of the accompanying figure (feedback calibration figure) clearly shows the adjustment process of this core parameter: the growth factor of the initial prediction Figure 2 (left column) is calibrated to (right column), and directly leads to the final employment prediction value from 875,000 to 798,000. Generate final prediction: substitute the calibrated parameter into

[0142]

[0143]

[0144] For example Figure 2 ​As shown, this value is simulated and verified as a new target, and the deviation D' of the new curve from the benchmark trend is reduced to 3.2%, which is lower than the threshold. Therefore, the system adopts = 798,000 people as the final prediction result. At this point, the complete closed-loop process shown in FIG. 1 is completed, and in particular, the feedback calibration loop (red dashed arrow) that points back to step S3 is implemented, achieving self-optimization of the model parameters. Figure 1

[0145] Step S6: Batch prediction

[0146] For other industries such as the mining industry, the construction industry, etc., steps S1 to S5 are repeated. The key difference is that each industry needs to use its own historical data in step S2 , to independently calibrate a set of model parameters specific to that industry , thereby constructing an industry-specific dynamic simulation twin model.

[0147] The above only describes preferred embodiments of the present application, and it should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be considered within the scope of protection of the present application.​

Claims

1. A method for predicting urban employment scale based on dynamic simulation and feedback calibration, characterized in that, Includes the following steps: S1. Data acquisition and processing: Acquire production capacity and employment data of the target industry in multiple historical census years, and calculate the historical comprehensive growth factor sequence based on the data; S2. Construct and calibrate a dynamic simulation twin model: Using system dynamics, establish a dynamic simulation twin model with production capacity and employment as state variables; based on historical production capacity and employment data and annual investment sequences estimated from historical data, determine an objective search space for model parameters, and calibrate the parameters through optimization algorithms to obtain a validated model that can reproduce historical dynamics. S3. Make initial forecasts: Forecast the industry capacity and comprehensive growth factor for the target year, and calculate the first employment forecast value based on the preset employment forecast equation. S4. Perform twin trend simulation: Use the first employment forecast value as a constraint on future employment, input it into the verified dynamic simulation twin model, and run the simulation to obtain simulated employment trend data for the target year and the subsequent preset period. S5. Trend Comparison and Deviation Analysis: Generate benchmark employment trend data based on the historical employment data, compare the simulated employment trend data obtained in step S4 with the benchmark employment trend data, and calculate the trend deviation between the two. S6. Feedback calibration and output: Determine whether the trend deviation exceeds a preset threshold; If so, then with reducing the deviation as the optimization objective, the parameters used to predict the comprehensive growth factor in step S3 are reverse-calibrated, and the comprehensive growth factor and employment scale are recalculated using the calibrated parameters to obtain the final employment prediction value and output it. If not, the first employment forecast value will be output directly as the final forecast result.

2. The method according to claim 1, characterized in that, In step S2, the core dynamic equations of the dynamic simulation twin model are as follows: in, and They represent the first Annual production capacity and employment stock; and They represent the first Annual fixed asset investment and comprehensive growth factors; The model parameters to be calibrated are, respectively, the investment conversion efficiency coefficient, the capacity-to-employment ratio, the growth factor influence intensity, the capacity depreciation rate, and the natural job loss rate.

3. The method according to claim 2, characterized in that, In step S2, determining the objective search space for the model parameters specifically includes: The capacity employment-driving coefficient The search scope is determined by calculating the instantaneous ratio of employment to production capacity for each historical census year. And based on the minimum and maximum values ​​of this ratio sequence, it is extended to determine; The investment conversion efficiency coefficient The search scope is determined by calculating efficiency estimates for multiple periods based on the increase in production capacity and the estimated total investment during the same period between adjacent historical census years, and then expanding the scope based on the minimum and maximum values ​​of the estimated value sequence. The capacity depreciation rate With natural job loss rate The search scope is determined based on the relevant national asset depreciation regulations and the reasonable scope published in the industry's labor mobility statistics report.

4. The method according to claim 2 or 3, characterized in that, In step S2, the parameter calibration through the optimization algorithm specifically involves: using the particle swarm optimization algorithm, with the fitness function being the weighted root mean square error between the census year state value output by the model simulation and the corresponding historical observation value, and iteratively optimizing within the determined parameter search space to obtain the optimal parameter set.

5. The method according to claim 1, characterized in that, In step S3, the first employment forecast value is calculated based on a preset employment forecast equation, and the equation used is: in, For the projected number of jobs in the target year n, For the projected production capacity in the target year n, The projected composite growth factor for the target year n; These represent the number of jobs and production capacity in the preceding historical census year, which is immediately adjacent to the target year.

6. The method according to claim 5, characterized in that, In step S3, the industry production capacity for the target year is... and comprehensive growth factors The predictions are achieved by performing nonlinear regression fitting on the corresponding historical data sequences.

7. The method according to claim 1 or 5, characterized in that, In step S6, the parameters used to predict the comprehensive growth factor are reverse-calibrated. Specifically, while keeping the form of the prediction model unchanged, the parameters of the prediction model are adjusted using a one-dimensional search or gradient descent method with the goal of minimizing the trend deviation, until the employment value calculated by the growth factor predicted using the new parameters is lower than the preset threshold.

8. The method according to claim 1, characterized in that, In step S5, the benchmark employment trend data is obtained by performing time series decomposition on the historical employment data series and extracting its long-term trend term, or by using exponential smoothing for fitting and extrapolation. The trend deviation is obtained by calculating the average absolute percentage error or root mean square error between the simulated employment trend data and the benchmark employment trend data within the same time interval.

9. The method according to claim 1, characterized in that, The method is executed in a batch processing manner, and the prediction described in claim 1 is performed independently for multiple different industries in turn. Each industry independently completes the model construction and calibration in step S2 using its own historical data.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 9.

Images