Human resource demand prediction method and system

Abstract

The invention relates to the technical field of human resources, in particular to a human resource demand prediction method and system, and the method comprises the steps: collecting internal data and external environment data of an enterprise, and obtaining standardized feature data after processing; a dual-channel weight calculation module constructed based on an attention mechanism respectively calculates a dependency weight of historical time sequence data and an influence weight of external environment data, and fuses the dependency weight and the influence weight to obtain a total weight of each data; based on the standardized feature data and the total weight of each data, a preset basic time sequence model is adopted to capture a periodic trend, an enhanced neural network is combined to learn a long-term and short-term dependency relationship, and meanwhile, a causal inference engine is constructed to quantify causal association between variables to obtain an initial demand prediction result; the dynamic error between the initial prediction result and the actual demand is calculated, when the error exceeds a preset adjustment threshold value, incremental learning is triggered to update the model parameters, and the calibrated demand prediction result is obtained, and the method has the advantages of being high in dynamic adaptability and high in prediction precision.

Description

Technical Field

[0001] This application relates to the technical field of human resources, and in particular to a method and system for forecasting human resource demand. Background Technology

[0002] In business operations and management, human resource demand forecasting is a core prerequisite for formulating recruitment plans and optimizing human resource allocation. Significant deviations in forecasting can lead to two extreme problems: first, overestimating demand results in over-recruiting, soaring labor costs, and redundant personnel reducing organizational efficiency; second, underestimating demand leads to vacancies in key positions, hindering business progress and even causing missed market opportunities. Therefore, accurate human resource demand forecasting is crucial for business development.

[0003] However, current mainstream human resource demand forecasting technologies still have significant shortcomings and are difficult to adapt to complex and ever-changing market environments. These shortcomings include: Traditional forecasting methods rely solely on historical data from within the company, such as the number of job openings, on-the-job rate, employee attendance records, and turnover rate over the past 3-5 years, while completely ignoring the impact of external market dynamics and real-time environmental factors.

[0004] Traditional forecasting models are mostly static, such as ARIMA and Markov chain models in time series analysis. The core logic of these models is to "infer future trends based on historical patterns", but they cannot cope with sudden and unconventional changes in demand. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method and system for forecasting human resource demand.

[0006] This application provides a method for forecasting human resource demand, which employs the following technical solution: A method for forecasting human resource demand includes the following steps: Collect internal enterprise data and external environmental data, process them to obtain standardized feature data; The dual-channel weight calculation module, built based on the attention mechanism, calculates the dependency weight of historical time series data and the influence weight of external environment data respectively, and then merges them to obtain the total weight of each data. Based on standardized feature data and the total weight of each data point, a pre-set basic time series model is used to capture periodic trends. Combined with an enhanced neural network to learn long-term and short-term dependencies, and at the same time, a causal inference engine is built to quantify the causal relationships between variables to obtain the initial demand prediction results. The dynamic error between the initial prediction result and the actual demand is calculated. When the error exceeds the preset adjustment threshold, incremental learning is triggered to update the model parameters and obtain the calibrated demand prediction result.

[0007] In one embodiment, the step of obtaining standardized feature data after processing specifically includes: An interpolation resampling algorithm is used to unify data of different frequencies into timestamps with a preset time granularity; Spatial encoding of data links internal enterprise data with external environmental data to the same spatial coordinate system; After spatiotemporal alignment, outlier detection algorithms are used to remove noise from the data, and a pre-defined semantic mapping table is used to unify the semantic standards of different data sources to obtain standardized feature data.

[0008] In one embodiment: the dual-channel weight calculation module includes a temporal attention submodule and an environmental attention submodule; the step of constructing the dual-channel weight calculation module based on the attention mechanism, calculating the dependency weight of historical time-series data and the influence weight of external environmental data respectively, and fusing them to obtain the total weight of each data specifically includes: Based on the temporal attention submodule, the historical temporal feature vector in the standardized feature data is used as input to calculate the dependence weight of the current moment on the historical data. Based on the environmental attention submodule, the influence weight of each environmental factor on the current demand is calculated by taking the external environmental feature vector in the standardized feature data as input. A gating function is used to fuse the temporal attention weights and the environmental attention weights to obtain the total weight of each data point.

[0009] In one embodiment: the step of calculating the dependency weight of the current moment on historical data based on the temporal attention submodule, using the historical temporal feature vector in the standardized feature data as input, specifically includes: Construct historical time-series feature vectors by dividing historical time-series data in standardized feature data according to time windows; A bidirectional LSTM layer is introduced into the temporal attention submodule. The historical temporal feature vector is used as input. The forward LSTM captures the short-term dependencies of historical data, and the backward LSTM captures the long-term dependencies. The output of the bidirectional LSTM is concatenated with the feature vector at the current time step and input into the fully connected layer to obtain the initial temporal weights. Then, the weights are normalized by the Softmax function to obtain the final temporal attention weights.

[0010] In one embodiment: after obtaining the standardized feature data, the process further includes: Based on the Pearson correlation coefficient, the correlation between each feature in the standardized feature data and human resource demand is calculated, and features with a correlation higher than the preset correlation threshold are selected as core features. For the core features after screening, multicollinearity is detected by variance inflation factor, and redundant features with variance inflation factor higher than the preset detection threshold are removed. The Random Forest algorithm was used to score the importance of the core features and sort them from high to low scores.

[0011] In one embodiment: the initial demand prediction result is obtained by using a preset basic time series model to capture periodic trends based on standardized feature data and the total weight of each data point, combined with an enhanced neural network to learn long-term and short-term dependencies, and simultaneously constructing a causal inference engine to quantify the causal relationships between variables. Specifically, this includes: Based on standardized feature data and the total weight of each data point, a hybrid temporal model is constructed using the Prophet model as the basic temporal module and the LSTM-Attention network as the enhancement module. Based on a pre-defined structural equation model, highly important features are selected from standardized feature data as core causal variables, the causal relationships between variables are quantified, and a causal inference engine is constructed. The output of the hybrid time series model is fused with the causal correction coefficients of the causal inference engine to obtain the initial demand prediction results.

[0012] In one embodiment: the training method of the causal inference engine includes: The training set is based on historical core causal variables and their corresponding human resource needs. By employing a causal loss function and training with a gradient descent optimizer, spurious correlations between variables are eliminated, ensuring that the causal coefficients output by the model match the actual business logic.

[0013] In one embodiment: after constructing the causal inference engine, the following is also included: Based on the pre-defined causal relationships between variables according to business logic, an initial causal path diagram is formed, and causal hypotheses are constructed. For each causal path in the causal hypothesis, perform do-calculus verification; The prediction results of the causal inference engine are compared with the actual human resource needs of similar historical scenarios. If the error exceeds the second preset error threshold, the causal hypothesis is corrected and the causal inference engine is retrained.

[0014] In one embodiment: the step of calculating the dynamic error between the initial prediction result and the actual demand, and triggering incremental learning to update the model parameters when the error exceeds a preset adjustment threshold, to obtain the calibrated demand prediction result specifically includes: By combining time decay weights, the deviation between the initial prediction results and the actual demand is evaluated to obtain the dynamic error; Based on a preset adjustment threshold, the dynamic error is judged, and if it exceeds the preset adjustment threshold, incremental learning is triggered. Based on dynamic error, the model parameters are dynamically adjusted to complete incremental learning.

[0015] This application provides a human resource demand forecasting system, which adopts the following technical solution: A human resource demand forecasting system includes the following steps: The data acquisition and standardization module is used to collect internal enterprise data and external environmental data, and process the collected data to obtain standardized feature data. The weight calculation module is a dual-channel weight calculation module built based on the attention mechanism. It calculates the dependency weight of historical time series data and the influence weight of external environment data respectively, and then merges them to obtain the total weight of each data. The prediction module, based on the standardized feature data obtained by the data acquisition and standardization module and the total weight of each data obtained by the weight calculation module, uses a preset basic time series model to capture periodic trends, combines an enhanced neural network to learn long-term and short-term dependencies, and simultaneously constructs a causal inference engine to quantify the causal relationship between variables to obtain the initial demand prediction result. The error calibration module calculates the dynamic error between the initial prediction result and the actual demand. When the error exceeds the preset adjustment threshold, it triggers incremental learning to update the model parameters and obtains the calibrated demand prediction result.

[0016] In summary, this application has the following beneficial effects: 1. Break through the limitations of relying on single data sources, achieve effective integration of internal enterprise data and external environmental data, and effectively improve the accuracy of prediction; 2. Enhance the dynamic adaptability of the model based on external environmental data, respond quickly to sudden events, calibrate prediction results in real time, and avoid lag. Attached Figure Description

[0017] Figure 1 This is a flowchart of a human resource demand forecasting method; Figure 2 This is a logical block diagram of a human resource demand forecasting system.

[0018] In the diagram, 10 is the data acquisition and standardization module; 20 is the weight calculation module; 30 is the prediction module; and 40 is the error calibration module. Detailed Implementation

[0019] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. In some cases, to avoid obscuring various aspects of this application due to unnecessary description, well-known methods, processes, systems, components, and / or circuits already described at a higher level will not be elaborated upon. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0020] It should be noted that the descriptions of these embodiments are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0021] In the description of this application, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0022] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples.

[0023] like Figure 1 As shown, this application discloses a method for forecasting human resource demand, including... S200: Collect internal enterprise data and external environmental data, and process them to obtain standardized feature data.

[0024] In this step, the internal data of the enterprise mainly includes enterprise business data, historical recruitment data, employee turnover data, performance data, and training data. Enterprise business data mainly refers to data that reflects business growth. Historical recruitment data includes data such as historical human resource demand, on-the-job rate, and recruitment cycle. Employee turnover data includes data such as turnover rate, retention rate, and reasons for leaving. Performance data includes data such as KPI completion rate and performance level distribution. Training data includes data such as skills improvement curve and training conversion rate.

[0025] External environment data mainly includes industry data, policy data, competitor data, and public opinion data. Among them, industry data includes talent mobility index and job demand popularity, policy data includes industry support policies and regional talent subsidies, competitor data includes competitor recruitment volume and salary range, and public opinion data includes job search volume on recruitment platforms and discussion popularity on professional social networking platforms.

[0026] Because internal enterprise data and external environmental data are not uniform in terms of source and format, they need to be aligned in time and space to form standardized feature data before they can be used.

[0027] In one embodiment, the spatiotemporal alignment step specifically includes: An interpolation resampling algorithm is used to unify data of different frequencies into timestamps with a preset time granularity; Geographic Information System (GIS) is used to spatially encode the data, linking internal enterprise data and external environmental data to the same spatial coordinate system; After spatiotemporal alignment, outlier detection algorithms such as the 3σ criterion are used to remove data noise, and a pre-defined semantic mapping table is used to unify the semantic standards of different data sources to obtain standardized feature data.

[0028] Among these methods, the interpolation resampling algorithm preferably aligns the time dimension of the data to the hourly or daily level. Spatial dimension alignment, on the other hand, is used to unify the data region. For example, if it is clear that job B is being recruited in city A, then the corresponding policy data and competitor data will both focus on city A.

[0029] Semantic standardization can be used to unify semantics and avoid data chaos caused by unclear semantic definitions. For example, the definition of "high-performing employee" in external industry reports can be mapped to the internal standard of "KPI completion rate ≥ 120%".

[0030] Furthermore, even after feature standardization, the problems of "weak data correlation and low contribution to prediction" may still exist. Therefore, further processing of the standardized feature data is necessary, specifically including: Feature correlation calculation: Based on the Pearson correlation coefficient, the correlation between each feature in the standardized feature data and human resource demand is calculated, and features with a correlation higher than the preset correlation threshold are selected as core features.

[0031] Feature redundancy removal: For the core features after screening, multicollinearity is detected by variance inflation factor, and redundant features with variance inflation factor higher than the preset detection threshold are removed. Feature importance ranking: The core features are scored using the random forest algorithm and ranked from highest to lowest score.

[0032] The training method for the above-mentioned random forest algorithm is as follows: The training set and validation set are based on the core historical features and the corresponding actual human resource needs. The training set accounts for 70%-80%, with an optimal 75%, and the validation set accounts for 20%-30%, with an optimal 25%.

[0033] The mean squared error loss function is used for iterative training until the stability of the feature importance ranking meets the requirements. In this embodiment, the condition for meeting the requirements is set as the ranking change rate being less than 5% for 10 consecutive iterations.

[0034] S400, a dual-channel weight calculation module based on the attention mechanism, calculates the dependency weight of historical time series data and the influence weight of external environment data respectively, and merges them to obtain the total weight of each data.

[0035] In one embodiment, the dual-channel weight calculation module includes a temporal attention submodule and an environmental attention submodule.

[0036] The training methods for the temporal attention submodule include: Using historical time-series data and corresponding human resource needs as training data, historical time-series data refers to a set of single-moment / continuous single-moment data that are strongly correlated with time in standardized feature data; The mean squared error loss function is adopted, and the SGD optimizer is trained to match the weight calculation results with the actual impact of historical time series data on the demand. The learning rate of the SGD optimizer ranges from 0.001 to 0.01, with 0.005 being preferred.

[0037] Training methods for the environmental attention submodule include: Historical external environment data and the corresponding changes in human resource demand were used as training data. The cross-entropy loss function is used, and the Adam optimizer is trained until the correlation coefficient between the environmental factor weights and demand changes exceeds a preset correlation threshold. The learning rate of the Adam optimizer ranges from 0.0005 to 0.002, with 0.001 being preferred.

[0038] Based on the above, step S400 specifically includes: Based on the temporal attention submodule, the historical temporal feature vector in the standardized feature data is used as input to calculate the dependence weight of the current moment on the historical data. Based on the environmental attention submodule, the influence weight of each environmental factor on the current demand is calculated by taking the external environmental feature vector in the standardized feature data as input. A gating function is used to fuse the temporal attention weights and the environmental attention weights to obtain the total weight of each data point.

[0039] To address the issue that temporal attention weights cannot accurately capture the long-term and short-term dependencies in historical data, the following steps, based on the environmental attention submodule, use the external environmental feature vector from standardized feature data as input to calculate the influence weights of each environmental factor on the current demand: Construct historical time-series feature vectors by dividing historical time-series data in standardized feature data according to time windows; A bidirectional LSTM layer is introduced into the temporal attention submodule. The historical temporal feature vector is used as input. The forward LSTM captures the short-term dependencies of historical data, and the backward LSTM captures the long-term dependencies. The output of the bidirectional LSTM is concatenated with the feature vector at the current time step and input into the fully connected layer to obtain the initial temporal weights. Then, the weights are normalized by the Softmax function to obtain the final temporal attention weights.

[0040] Training methods for bidirectional LSTM include: The training data consists of historical time-series feature vectors and corresponding actual human resource needs. A time-series prediction loss function is used, and the model is trained using the Adam optimizer until its accuracy in predicting dependencies on historical time-series data exceeds a first preset accuracy threshold. The learning rate of the Adam optimizer ranges from 0.0005 to 0.0015, with 0.001 being preferred.

[0041] The formula for calculating the dependency weight of the current moment on historical data is as follows: In the formula, H t H represents the time-series feature vector for the current moment, containing historical related data such as recruitment demand and turnover rate at that moment; t-1 W is the temporal feature vector of the previous time step; a The weight matrix learned by the model during training is used to adjust H. t With H t-1 The importance of b; aThe bias term is used to correct calculation errors; softmax is the normalization function to ensure that α... t The value of α is between 0 and 1, and the α value at all times is... t The sum is 1.

[0042] The formula for calculating the weight of each environmental factor on current demand is as follows: In the formula, E is an environmental factor vector, containing data such as current policy data, competitor data, and public opinion data; W b This is a weight matrix for environmental factors, used to adjust the importance of different environmental factors; b b This is a bias term.

[0043] The gating function is the Sigmoid function, and the final weight of each data point is calculated using the following formula: In the formula, σ is the Sigmoid function, which controls the fused weights between 0 and 1 to avoid imbalanced predictions caused by excessively large or small weights; ⊗ represents element-wise multiplication, i.e., α t β is multiplied by the corresponding elements of W1, and β is multiplied by the corresponding elements of W2; W1 and W2 are the basic parameter matrices obtained from model training, used to balance the initial contributions of time series data and environmental data.

[0044] S600 uses standardized feature data and the total weight of each data point to capture periodic trends using a preset basic time series model. It combines this with an enhanced neural network to learn long-term and short-term dependencies and simultaneously builds a causal inference engine to quantify the causal relationships between variables, thus obtaining the initial demand prediction results.

[0045] To address the aforementioned issues of poor robustness in single-model predictions and the inability to distinguish between correlation and causation leading to prediction bias, step S600 specifically includes: Based on standardized feature data and the total weight of each data point, a hybrid temporal model is constructed using the Prophet model as the basic temporal module and the LSTM-Attention network as the enhancement module. Based on a pre-defined structural equation model, highly important features are selected from standardized feature data as core causal variables, the causal relationships between variables are quantified, and a causal inference engine is constructed. The output of the hybrid time series model is fused with the causal correction coefficients of the causal inference engine to obtain the initial demand prediction results.

[0046] The training method for the Prophet model includes: using historical human resource demand from standardized feature data as the training set; optimizing the trend and seasonal parameters of the model through maximum likelihood estimation, so that the model's fitting error to historical periodic trends is lower than a first preset error threshold.

[0047] The training method for the LSTM-Attention network includes: using historical time-series feature vectors and corresponding actual human resource needs as training data; employing the mean squared error loss function and training with the Adam optimizer until the model's prediction accuracy for long-term and short-term dependencies exceeds a second preset accuracy threshold. The learning rate of the Adam optimizer ranges from 0.0001 to 0.001, with 0.0005 being preferred.

[0048] The training method for the causal inference engine includes: using historical core causal variables and corresponding human resource needs as the training set; employing a causal loss function and training through a gradient descent optimizer to eliminate spurious correlations between variables, so that the causal coefficients output by the model match the actual business logic.

[0049] Among them, high-importance features are obtained based on importance scores, and features with importance scores higher than a preset score threshold are used as core causal variables.

[0050] In one embodiment, to address the problem that the constructed causal reasoning engine cannot effectively verify the accuracy of causal relationships and is prone to retaining spurious associations, the causal reasoning engine further includes the following after its construction: Logically presuppose the causal direction between variables, form an initial causal path diagram, and construct causal hypotheses; For each causal path in the causal hypothesis, perform do-calculus verification; If the error between the prediction results of the causal inference engine and the actual human resource needs of similar historical scenarios exceeds the second preset error threshold, the causal hypothesis is corrected and the causal inference engine is retrained.

[0051] Specifically, during the do-calculus verification, based on each causal path, a certain variable value is fixed through do-calculus, and the change in requirements is observed to verify the causal validity of the causal path; if the deviation between the change and the preset business logic requirement change exceeds the preset deviation threshold, the causal coefficient of the path is adjusted.

[0052] Among them, the preset business logic is the common sense of cause and effect summarized by enterprises or industries based on long-term operational practices and historical data. It serves as the business baseline for judging the rationality of the model, and the change in business logic requirements is obtained based on the summary of historical data.

[0053] S800: Calculate the dynamic error between the initial prediction result and the actual demand. When the error exceeds the preset adjustment threshold, trigger incremental learning to update the model parameters and obtain the calibrated demand prediction result.

[0054] Specifically, step S800 includes: The dynamic mean square error (DMAE) is used, combined with time decay weights, to evaluate the deviation between the initial prediction results and the actual requirements, and obtain the dynamic error. Based on a preset adjustment threshold, the dynamic error is judged, and if it exceeds the preset adjustment threshold, incremental learning is triggered. Based on dynamic error, the model parameters are dynamically adjusted to complete incremental learning.

[0055] The parameters of the aforementioned time decay weights are determined through training. The training data consists of historical error data and the corresponding model adjustment effects. The goal is to make the contribution of recent errors to the weights higher than that of long-term errors.

[0056] The formula for calculating the dynamic mean square error is: In the formula, N is the number of samples used for error calculation, such as prediction data from the past 72 hours; y i This represents the actual human resource demand at time i. For the predicted demand at time i; w i γ is the time decay weight at time i; γ is the decay coefficient, defaulting to 0.1; t now t represents the current time; i Let i be the i-th time.

[0057] The formula for incremental learning is: In the formula, For the updated model parameters, such as the weight matrix W of the attention module a Coefficients of the mixed model, etc.; The parameters are the model parameters before the update; λ is the decay coefficient, which defaults to 0.05 and is used to control the magnitude of parameter updates to avoid the model "overcorrecting" due to excessive single errors.

[0058] like Figure 2 As shown, this application also discloses a human resource demand forecasting system, including the following steps: The data acquisition and standardization module 10 is used to collect internal enterprise data and external environmental data, and process the collected data to obtain standardized feature data. The weight acquisition module 20 is a dual-channel weight calculation module built based on the attention mechanism. It calculates the dependency weight of historical time series data and the influence weight of external environment data respectively, and then merges them to obtain the total weight of each data. The prediction module 30, based on the standardized feature data obtained by the data acquisition and standardization module 10 and the total weight of each data obtained by the weight acquisition module 20, uses a preset basic time series model to capture periodic trends, combines an enhanced neural network to learn long-term and short-term dependencies, and simultaneously constructs a causal inference engine to quantify the causal relationship between variables, thereby obtaining the initial demand prediction result. The error calibration module 40 calculates the dynamic error between the initial prediction result and the actual demand. When the error exceeds the preset adjustment threshold, it triggers incremental learning to update the model parameters and obtains the calibrated demand prediction result.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for forecasting human resource demand, characterized in that, Includes the following steps: Collect internal enterprise data and external environmental data, process them to obtain standardized feature data; The dual-channel weight calculation module, built based on the attention mechanism, calculates the dependency weight of historical time series data and the influence weight of external environment data respectively, and then merges them to obtain the total weight of each data. Based on standardized feature data and the total weight of each data point, a pre-set basic time series model is used to capture periodic trends. Combined with an enhanced neural network to learn long-term and short-term dependencies, and at the same time, a causal inference engine is built to quantify the causal relationships between variables to obtain the initial demand prediction results. The dynamic error between the initial prediction result and the actual demand is calculated. When the error exceeds the preset adjustment threshold, incremental learning is triggered to update the model parameters and obtain the calibrated demand prediction result.

2. The human resource demand forecasting method according to claim 1, characterized in that, The steps for obtaining standardized feature data after processing specifically include: An interpolation resampling algorithm is used to unify data of different frequencies into timestamps with a preset time granularity; Spatial encoding of data links internal enterprise data with external environmental data to the same spatial coordinate system; After spatiotemporal alignment, outlier detection algorithms are used to remove noise from the data, and a pre-defined semantic mapping table is used to unify the semantic standards of different data sources to obtain standardized feature data.

3. The human resource demand forecasting method according to claim 1, characterized in that, The dual-channel weight calculation module includes a temporal attention submodule and an environmental attention submodule; the specific steps of constructing the dual-channel weight calculation module based on the attention mechanism, calculating the dependency weight of historical time-series data and the influence weight of external environmental data respectively, and fusing them to obtain the total weight of each data include: Based on the temporal attention submodule, the historical temporal feature vector in the standardized feature data is used as input to calculate the dependence weight of the current moment on the historical data. Based on the environmental attention submodule, the influence weight of each environmental factor on the current demand is calculated by taking the external environmental feature vector in the standardized feature data as input. A gating function is used to fuse the temporal attention weights and the environmental attention weights to obtain the total weight of each data point.

4. The human resource demand forecasting method according to claim 3, characterized in that, The step of calculating the dependency weight of the current moment on historical data based on the time-series attention submodule, using the historical time-series feature vector in the standardized feature data as input, specifically includes: Construct historical time-series feature vectors by dividing historical time-series data in standardized feature data according to time windows; A bidirectional LSTM layer is introduced into the temporal attention submodule. The historical temporal feature vector is used as input. The forward LSTM captures the short-term dependencies of historical data, and the backward LSTM captures the long-term dependencies. The output of the bidirectional LSTM is concatenated with the feature vector at the current time step and input into the fully connected layer to obtain the initial temporal weights. Then, the weights are normalized by the Softmax function to obtain the final temporal attention weights.

5. The human resource demand forecasting method according to claim 1, characterized in that, After obtaining the standardized feature data, the following is also included: Based on the Pearson correlation coefficient, the correlation between each feature in the standardized feature data and human resource demand is calculated, and features with a correlation higher than the preset correlation threshold are selected as core features. For the core features after screening, multicollinearity is detected by variance inflation factor, and redundant features with variance inflation factor higher than the preset detection threshold are removed. The Random Forest algorithm was used to score the importance of the core features and sort them from high to low scores.

6. The human resource demand forecasting method according to claim 5, characterized in that, The process involves using standardized feature data and the total weight of each data point, employing a pre-defined basic time-series model to capture periodic trends, combining this with an enhanced neural network to learn long-term and short-term dependencies, and simultaneously constructing a causal inference engine to quantify causal relationships between variables. The resulting initial demand prediction specifically includes: Based on standardized feature data and the total weight of each data point, a hybrid temporal model is constructed using the Prophet model as the basic temporal module and the LSTM-Attention network as the enhancement module. Based on a pre-defined structural equation model, highly important features are selected from standardized feature data as core causal variables, the causal relationships between variables are quantified, and a causal inference engine is constructed. The output of the hybrid time series model is fused with the causal correction coefficients of the causal inference engine to obtain the initial demand prediction results.

7. The human resource demand forecasting method according to claim 6, characterized in that, The training method for the causal reasoning engine includes: The training set is based on historical core causal variables and their corresponding human resource needs. By employing a causal loss function and training with a gradient descent optimizer, spurious correlations between variables are eliminated, ensuring that the causal coefficients output by the model match the actual business logic.

8. The human resource demand forecasting method according to claim 7, characterized in that, The causal reasoning engine also includes the following after its construction: Based on the pre-defined causal relationships between variables according to business logic, an initial causal path diagram is formed, and causal hypotheses are constructed. For each causal path in the causal hypothesis, perform do-calculus verification; The prediction results of the causal inference engine are compared with the actual human resource needs of similar historical scenarios. If the error exceeds the second preset error threshold, the causal hypothesis is corrected and the causal inference engine is retrained.

9. The human resource demand forecasting method according to claim 1, characterized in that, The steps of calculating the dynamic error between the initial prediction result and the actual demand, and triggering incremental learning to update the model parameters when the error exceeds a preset adjustment threshold to obtain the calibrated demand prediction result, specifically include: By combining time decay weights, the deviation between the initial prediction results and the actual demand is evaluated to obtain the dynamic error; Based on a preset adjustment threshold, the dynamic error is judged, and if it exceeds the preset adjustment threshold, incremental learning is triggered. Based on dynamic error, the model parameters are dynamically adjusted to complete incremental learning.

10. A human resource demand forecasting system, characterized in that, include: The data collection and standardization module (10) is used to collect internal enterprise data and external environment data, and to process the collected data to obtain standardized feature data. The weight acquisition module (20) constructs a dual-channel weight calculation module based on the attention mechanism, calculates the dependency weight of historical time series data and the influence weight of external environment data respectively, and then integrates them to obtain the total weight of each data. The prediction module (30) uses the standardized feature data obtained by the data acquisition and standardization module (10) and the total weight of each data obtained by the weight acquisition module (20) to capture periodic trends using a preset basic time series model, and combines the enhanced neural network to learn the long-term and short-term dependencies. At the same time, it constructs a causal reasoning engine to quantify the causal relationship between variables and obtains the initial demand prediction result. The error calibration module (40) calculates the dynamic error between the initial prediction result and the actual demand. When the error exceeds the preset adjustment threshold, it triggers incremental learning to update the model parameters and obtains the calibrated demand prediction result.

Images