Vocational education demand prediction system and dynamic adjustment method based on LSTM (Long Short Term Memory) time sequence prediction

Abstract

The invention relates to the technical field of vocational education, in particular to a vocational education demand prediction system based on LSTM time sequence prediction and a dynamic adjustment method, and the system comprises a multi-source data collection module, a knowledge graph construction module, an intelligent data processing module, an LSTM-reinforcement learning prediction module, a dynamic adjustment module and an execution feedback module. According to the scheme, firstly, three types of data of industrial economy, post requirements and education and learning are integrated, a knowledge graph containing semantic and quantitative information is constructed, abnormal value recognition and missing value filling are optimized, and data suitability is improved; secondly, capturing time sequence dependence through bidirectional LSTM, optimizing model parameters in combination with knowledge attention and reinforcement learning, and achieving demand prediction; and finally, constructing a resource dynamic coefficient matrix, converting constraints into mathematical conditions to solve an optimal adjustment amount, and feeding back an optimization model and a map in real time, thereby effectively solving the problems of extensive data, prediction deviation and experience-dependent adjustment in the prior art, and realizing efficient matching of vocational education resources and demands.

Description

Technical Field

[0001] This invention relates to the field of vocational education technology, and in particular to a vocational education demand forecasting system and dynamic adjustment method based on LSTM time series forecasting. Background Technology

[0002] Currently, with the accelerated upgrading of industrial structure and the transformation of economic model, vocational education, as a key link in ensuring the supply of human resources and matching industrial needs, is increasingly making the scientific and timely prediction of demand a core demand for industry development.

[0003] While existing vocational education demand forecasting schemes have improved the rationality of demand assessment to some extent, they still suffer from significant technical deficiencies: First, the data processing mechanism is crude, with outlier detection relying heavily on conventional fixed standards, failing to differentiate between real industry fluctuations and data errors by incorporating dynamic industry characteristics. Missing value imputation often employs simple time-series interpolation or mean-filling methods, ignoring the inherent logical relationships between entities, resulting in data quality that cannot match the dynamic changes in industry demand. Second, the forecasting model lacks adaptability. Traditional LSTM models only focus on the time-series dependencies of historical data, lacking mechanisms to cope with special situations such as adjustments to industry policies and sudden industry changes. Furthermore, they fail to optimize the weights of key elements such as key industries and in-demand skills, making them prone to forecasting bias. Third, dynamic adjustment and feedback loops are lacking. Most schemes only output demand forecast results, failing to translate industry linkage requirements and school resource constraints into quantifiable resource adjustment criteria. Adjustment scheme formulation relies heavily on manual experience, and there is a lack of real-time monitoring of the execution process and model iteration optimization mechanisms, leading to a disconnect between vocational education resource allocation and demand changes. Therefore, we propose a vocational education demand forecasting system and dynamic adjustment method based on LSTM time-series forecasting. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a vocational education demand forecasting system and dynamic adjustment method based on LSTM time series forecasting, thereby solving the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The vocational education demand prediction system based on LSTM time series prediction includes a multi-source data acquisition module, a knowledge graph construction module, an intelligent data processing module, an LSTM-reinforcement learning prediction module, a dynamic adjustment module, and an execution feedback module.

[0007] The multi-source data acquisition module connects to government industry databases, recruitment information platforms, and vocational school management systems through customized interfaces to collect three types of data: industry economic indicators, job demand information, and education and school operation data.

[0008] The knowledge graph construction module uses natural language processing technology to extract core entities from text data, combines rules in the field of vocational education to build entity relationships, calculates the association strength through entity co-occurrence frequency, and binds numerical data as attributes to corresponding entities to form a knowledge system containing semantic associations and quantitative information.

[0009] The intelligent data processing module combines knowledge graph association rules to identify and correct time-series outliers, fills data gaps with strong association entity features, and generates high-order features based on entity association paths.

[0010] The LSTM-reinforcement learning prediction module uses a bidirectional LSTM network to capture the temporal dependence of data, introduces a knowledge attention mechanism to strengthen the feature weights of key elements, and dynamically optimizes the model parameters by combining reinforcement learning algorithms. It outputs the prediction results of the quantity and structure of demand through multi-task learning and calibration.

[0011] The dynamic adjustment module constructs a dynamic coefficient matrix for enrollment plans, curriculum design, and faculty allocation based on the importance weights of entities in the knowledge graph. It transforms industry linkage constraints and school resource limits into mathematical constraints and solves the adjustment amount that maximizes resource utilization through optimization algorithms.

[0012] The execution feedback module decomposes and adjusts the tasks into specific tasks based on the relationship between the responsible entities and tasks in the knowledge graph and clarifies the execution standards. It monitors the execution data in real time, calculates the deviation rate, and feeds the deviation information back to the front-end module to optimize the model parameters and knowledge association rules.

[0013] In one possible implementation, the multi-source data acquisition module collects industrial economic indicators including the monthly output growth rate of key industries, the proportion of industrial structure, and the regional unemployment rate. These indicators are obtained periodically through the government open platform API interface, with the collection cycle dynamically set according to the data granularity. Job demand information covers the monthly recruitment volume, skill set, and average salary for job positions. Distributed web crawlers are used to collect data from mainstream recruitment platforms and filter invalid positions without clear skill requirements, and the industry to which the positions belong is marked. Educational data includes the number of students enrolled in each major, the graduate employment rate, and the core curriculum settings. This data is obtained in an anonymized manner, retaining only the summary data at the major level.

[0014] In one possible implementation, the knowledge graph construction module uses a BERT-based pre-trained model combined with an occupational classification dictionary to extract four types of entities: industry, job, major, and skill. Based on entity co-occurrence frequency and domain rules, it constructs three types of association relationships: industry includes job, job requires skills, and major requires skills. The association strength is calculated and normalized to the 0-1 range. Numerical data such as job recruitment volume, salary, major enrollment, and employment rate are bound to the corresponding job entity and major entity, respectively.

[0015] In one possible implementation, when the intelligent data processing module identifies time-series outliers, it uses a local weighted regression algorithm to fit the trend curve, constructs a dynamic threshold based on industry volatility, and verifies the consistency between the data and the trends of related entities through a knowledge graph. If there is a discrepancy, it uses the average value of the same industry and similar positions for correction. When filling data gaps, random gaps are filled using a method that combines time-series interpolation with the growth rate of strongly related entities, while systematic gaps are filled using a 2-layer GraphSAGE algorithm to aggregate and predict the features of related entities. When generating higher-order features, it calculates the path strength based on the entity association path, constructs industry-professional association features, skill demand features, and school quality features, filters features through mutual information entropy, and performs knowledge-weighted interval standardization.

[0016] In one possible implementation, a dynamic adjustment method for vocational education demand based on LSTM time-series forecasting includes the following steps:

[0017] S1: Multi-source data collection and knowledge graph construction, cross-channel aggregation of industrial economic data, job demand data, and vocational education data, extraction of core entities and construction of relationships, forming a system that integrates knowledge and data;

[0018] S2: Intelligent data processing, based on a fusion system, identifies and corrects outliers, fills in missing data, generates and filters high-order features, and outputs well-organized data;

[0019] S3: LSTM-Reinforcement Learning Fusion Prediction inputs regularized data into the fusion model, concatenates data features and entity embedding vectors, captures temporal dependencies through bidirectional LSTM, introduces knowledge attention mechanism and reinforcement learning to optimize the model, and outputs calibrated demand prediction results.

[0020] S4: Generate a dynamic adjustment plan, calculate the change in demand based on the forecast results and correct it through the industrial correlation, construct a dynamic coefficient matrix, and solve the optimal resource adjustment amount in combination with the constraints.

[0021] S5: Execution feedback and optimization, decompose the adjustment amount into specific tasks and execute them, monitor the execution process in real time to calculate the deviation rate, and provide feedback on deviation information to optimize the model and knowledge graph.

[0022] In one possible implementation, during step S1, monthly data is collected on a fixed date each month, and quarterly data is collected on a fixed date in the first month, with the statistical range and time stamp recorded synchronously. During knowledge graph construction, the entity association strength calculation formula is normalized to a value range of 0-1, thus integrating knowledge and data into a system. ,in, For knowledge graphs, , , These are data matrices for industrial economics, job demand, and vocational education.

[0023] In one possible implementation, in step S2, a fixed smoothing parameter is used when fitting the trend curve, and a Gaussian kernel function with a specific bandwidth is selected for weight calculation to balance trend capture and short-term fluctuations. The dynamic threshold is determined based on the industry's output volatility over the past six months. For random missing value imputation, a preliminary estimate is obtained by calculating the month-on-month growth rate using data from previous and subsequent periods, and then a corrected value is obtained by combining the growth rates of entities with sufficient correlation strength. The final imputation value is generated by weighted aggregation. For systematic missing value imputation, entities are treated as nodes, historical data and related entity attributes are used as features, and correlation strength is used as edge weights. The predicted imputation value is output by aggregating neighbor node information and inputting it into a fully connected layer. In the feature standardization stage, different numerical ranges are defined for key industry correlation features, scarce skill features, and ordinary features. Specifically, key industry correlation features and scarce skill features are standardized using... Intervals, other features are used The interval is calculated, and finally, regularized data is output.

[0024] In one possible implementation, in step S3, entity embeddings are generated using the TransE algorithm, and the input vector dimension is the sum of the data feature dimension and the entity embedding dimension. In the bidirectional LSTM, the forward LSTM learns historical trend features by forward propagation from the start time step to the end time step, and the hidden layer is initialized using Xavier. The backward LSTM integrates look-ahead information by backward propagation from the end time step to the start time step. The hidden states output by both are concatenated into a vector with a fixed number of hidden units. The knowledge attention mechanism assigns high weights to key entities and calculates attention weights by calculating the cosine similarity between the hidden state and the entity embedding vector, thus obtaining attention-enhanced hidden states. Reinforcement learning uses the DDPG algorithm to construct the agent. The state includes the preceding prediction bias, bias trend, and data update amount. The actions include the adjustment of the number of LSTM hidden units, learning rate, and attention weight coefficients. The reward function simultaneously penalizes prediction bias and drastic parameter changes. In the multi-task output layer, the quantity prediction branch is a 2-layer fully connected network, and the structure prediction branch is a combination of a 1-layer convolutional layer and a 2-layer fully connected layer. The multi-task loss function is used for optimization, and prediction values ​​exceeding the threshold are corrected based on knowledge graph rules.

[0025] In one possible implementation, step S4 includes changes in demand such as month-on-month changes, year-on-year changes, and trend changes, which are corrected by the average industrial correlation of the past 12 months in the knowledge graph. In the dynamic coefficient matrix, the enrollment coefficient matrix is ​​a diagonal matrix, and the diagonal elements are calculated by combining the proportion of industrial output value, professional employment rate, and industrial importance weight, with key industries having a higher weight than general industries. The course coefficient is calculated by combining the urgency of skill demand, the matching degree between existing courses and skills, and the importance weight of skills, with scarce skills having a higher weight. The teacher coefficient is calculated by combining the teacher-student ratio, teacher skill proficiency, and the importance weight of industry and skills. The objective function of the constraint optimization is to maximize resource utilization, and the constraints include the upper limit of enrollment plan, course adjustment cycle, and teacher staffing limit. The constraints are incorporated into the objective function using the Lagrange multiplier method, and the optimal adjustment amount is solved by combining the KKT conditions.

[0026] In one possible implementation, in step S5, the execution path and quantitative acceptance criteria are clearly defined during task decomposition. Tasks are pushed through the school's office management system, which supports progress reporting and feedback on questions. Execution difficulties are addressed within a specified time. Execution monitoring collects multi-dimensional data through the campus card system, training equipment management system, and teacher assessment system. The deviation rate is calculated by comparing the data with the execution standard entities in the knowledge graph. When the deviation rate exceeds a threshold, an early warning is triggered and corrective suggestions are pushed. During system optimization, the deviation rate is input into the reinforcement learning agent to adjust model parameters. The association strength of the knowledge graph is updated according to the cause of the deviation. The predicted MAE and resource matching rate before and after optimization are compared quarterly to verify the optimization effect.

[0027] Beneficial effects compared to existing technologies:

[0028] 1. This solution integrates three core data categories: industry economic indicators, job demand information, and educational data. It employs natural language processing technology combined with rules specific to vocational education to extract core entities such as industry, job title, major, and skills from text data. Through entity association analysis, a knowledge graph containing semantic relationships and quantitative information is constructed, forming a structured domain knowledge system. During data processing, outliers are identified by combining knowledge graph association rules with industry dynamic characteristics. Data confirmed as erroneous is corrected using similar references. For missing data, features of strongly related entities are used to fill in the gaps. This effectively addresses the problems of existing technologies' coarse data processing and lack of deep integration with domain knowledge, significantly improving the data's adaptability to changing demands.

[0029] 2. In this scheme, a bidirectional LSTM network is used to capture the dependencies between time-series data. Forward propagation learns historical trend features, while backward propagation integrates forward-looking information such as industry policies, effectively alleviating the gradient vanishing problem in long-sequence training and ensuring comprehensive capture of time-series patterns. Simultaneously, a knowledge attention mechanism is introduced, calculating the correlation between hidden states and entity embedding vectors to assign high weights to key entities such as key industries and in-demand skills, strengthening the influence of key elements on prediction results. Furthermore, a reinforcement learning algorithm is used to construct an intelligent agent, dynamically optimizing model parameters with prediction bias, bias trends, and data update characteristics as states. This ensures the model maintains stable prediction performance even when data distribution changes, enabling the prediction of the quantity and structure of vocational education demand.

[0030] 3. In this scheme, by constructing a knowledge-constrained quantitative dynamic adjustment and closed-loop feedback mechanism, the optimal allocation of vocational education resources and continuous iterative optimization of the system are achieved. Based on the importance weights of entities in the knowledge graph, a dynamic coefficient matrix of three types of resources—enrollment plan, curriculum design, and teacher allocation—is constructed. Practical requirements such as industry linkage constraints and school resource limits are transformed into mathematical constraints. The adjustment amount that maximizes resource utilization is solved through optimization algorithms, ensuring that the adjustment scheme achieves optimal allocation within the allowable range of school conditions, and avoiding the subjectivity and limitations of manually formulated schemes. Attached Figure Description

[0031] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0032] Figure 1 This is a schematic diagram of the system framework of the vocational education demand forecasting system of the present invention;

[0033] Figure 2 This is a flowchart illustrating the method for dynamically adjusting vocational education needs according to the present invention. Detailed Implementation

[0034] Preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the present invention can also be implemented in various different forms, and therefore the present invention is not limited to the embodiments described below. In addition, for the purpose of more clearly describing the present invention, parts not connected to the invention will be omitted from the drawings.

[0035] The technical solutions in this application are designed to address the problems described in the background, and are generally as follows:

[0036] Example:

[0037] Please refer to Figure 1As shown in Figure 2, this embodiment introduces a vocational education demand prediction system based on LSTM time series prediction, including a multi-source data acquisition module, a knowledge graph construction module, an intelligent data processing module, an LSTM-reinforcement learning prediction module, a dynamic adjustment module, and an execution feedback module.

[0038] The multi-source data acquisition module is responsible for basic data aggregation. This module connects to government industry databases, recruitment information platforms, and vocational school management systems through customized interfaces to collect three types of core data: industry economic indicators (such as monthly output growth rate and industry structure share), job demand information (such as recruitment quantity, skill requirements, and salary range), and education data (such as professional enrollment scale, graduate employment rate, and curriculum).

[0039] The knowledge graph construction module enables the structured integration of domain knowledge. This module employs natural language processing technology to extract core entities such as industries, job positions, majors, and skills from text data, and constructs entity relationships by combining them with rules from the vocational education domain. The module calculates the association strength (normalized to the 0-1 range) through entity co-occurrence frequency, and binds numerical data (such as job postings and professional employment rates) as attributes to corresponding entities, forming a knowledge system that includes semantic associations and quantitative information.

[0040] The intelligent data processing module performs deep normalization on the raw data. First, it identifies outliers in the time series by combining knowledge graph association rules, and corrects confirmed errors using the average of similar professional data over the past three months. To address missing data, it uses the features of strongly related entities to fill in the gaps. High-order features are generated based on entity association paths, thereby improving the data's ability to represent changes in demand.

[0041] The LSTM-reinforcement learning prediction module constructs a fusion model and obtains predictions. This module uses a bidirectional LSTM network to capture the temporal dependencies of the data, learns historical trend features through forward propagation, and integrates forward-looking information such as industrial policies through backpropagation. It introduces a knowledge attention mechanism to strengthen the feature weights of key elements such as key industries and scarce skills through similarity calculation of entity embedding vectors. It dynamically optimizes the model parameters by combining reinforcement learning algorithms, and outputs the prediction results of demand quantity and structure through multi-task learning. It is also calibrated based on knowledge rules such as job demand must not exceed the carrying capacity of the corresponding industry.

[0042] The dynamic adjustment module generates resource optimization schemes based on prediction results. This module constructs dynamic coefficient matrices for three types of resources—enrollment plans, curriculum design, and faculty allocation—based on the importance weights of entities in the knowledge graph. It transforms industry linkage constraints and educational resource limits into mathematical constraints, and uses optimization algorithms to solve for the adjustment amount that maximizes resource utilization, ensuring that the scheme achieves optimal allocation within the limits of available educational resources.

[0043] The execution feedback module is used to implement the solution and iteratively optimize the system. Based on the relationships between responsible entities and tasks in the knowledge graph, this module breaks down adjustments into specific tasks and clarifies execution standards. It also calculates the deviation rate by monitoring execution data in real time, feeding this deviation information back to the front-end module to dynamically optimize model parameters and knowledge association rules, continuously improving the system's adaptability to changing vocational education needs.

[0044] Based on the above system, this embodiment also introduces a dynamic adjustment method for vocational education demand based on LSTM time series prediction, which is used in conjunction with the above system. The specific steps are as follows:

[0045] S1: Multi-source data acquisition and knowledge graph construction

[0046] This step provides the foundational support for the system to integrate knowledge and data through cross-channel data aggregation and semantic association modeling. Data collection covers three core types of information: industrial economic data, including the monthly output growth rate of 15 key industries. Industrial structure proportion ( Regional unemployment rate Data is collected periodically via the government open platform API. The collection cycle is dynamically set according to the data granularity. For example, monthly data is collected on the 8th of each month, and quarterly data is collected on the 12th of the first month. The statistical scope (such as whether it includes county-level economy) and time stamp are recorded synchronously. Next, the job demand data covers the monthly recruitment volume for over 200 job positions. Skills Requirements Average salary ( The system uses distributed web crawlers to collect data from five major recruitment platforms. Natural language processing is used to filter out invalid job postings without specific skill requirements, and the industry to which the job postings belong is ultimately labeled. In addition, the vocational education data includes enrollment figures for over 80 majors. Graduate employment rate Core Curriculum ( The data was obtained from the provincial vocational education management system and anonymized. Sensitive information such as student ID numbers and home addresses were removed, and only summary data at the professional level was retained.

[0047] Knowledge graph construction achieves structured integration of domain knowledge through entity extraction and association modeling. It employs a BERT-based pre-trained model combined with an occupational classification dictionary to extract industry information from textual data. ,post ,major ,Skill Four types of core entities. Relationships are constructed based on entity co-occurrence frequency and domain rules, primarily including three core relationships: "industry includes job positions," "job positions require skills," and "professional training skills," with the strength of these relationships... The calculation formula is The normalized value range is Numerical data is bound as attributes to the corresponding entity; for example, the job title entity. Related recruitment volume With salary professional entities Related enrollment numbers With employment rate Ultimately, this will form a system that integrates knowledge and data. ,in, For knowledge graphs, , , These are data matrices for industrial economics, job demand, and vocational education.

[0048] S2: Intelligent Data Processing

[0049] This step is based on the fusion system. Deep data normalization is performed to improve data quality and feature representation capabilities through knowledge-guided learning. Anomaly handling is combined with knowledge graph association rules: for each time-series data point... The trend curve was fitted using a locally weighted regression algorithm. The smoothing parameter was set to 0.65, and the weights were calculated using a Gaussian kernel function. (Where the bandwidth is set to 5 to balance trend capture and short-term fluctuations), calculate the deviation rate. Constructing dynamic thresholds based on industry volatility ,in, For industry Output volatility over the past 6 months, when > In this case, the consistency between data and related entities' trends is verified using a knowledge graph. For example, abnormal fluctuations in the recruitment volume for a certain auto repair job need to be consistent with the growth trend of the auto industry. If they are consistent, it is determined to be a genuine fluctuation; if they are inconsistent, the average of similar jobs in the same industry over the past three months is used for correction. The correction formula is as follows: .

[0050] Next, entity association information is used to fill in the missing data. For random missing data (such as a gap in e-commerce job demand data for a certain month), a fusion method of time-series interpolation and knowledge matching is adopted: the time dimension calculates the month-on-month growth rate using data from the three months before and after. A preliminary estimate was obtained. Spatial dimension combined with strongly related entities in the knowledge graph (relationship strength) growth rate >0.8 , obtain the correction value The final fill value is This approach balances temporal continuity with the influence of related entities. For systematically missing data, a two-layer GraphSAGE algorithm is used to aggregate and predict related entity features. Entities are treated as nodes, features are historical data and related entity attributes, and edge weights represent the association strength. and through Aggregate neighbor information ( Embed vectors for nodes. For the set of neighboring nodes, (Using initial features) as input to the fully connected layer to output predicted values, so that the filling results fit the actual development trend of the associated entities.

[0051] Furthermore, the predictive power of the data is improved by constructing high-order semantic features. Path strength is calculated based on entity association paths in the knowledge graph. Reflecting the industry With professional The strength of indirect association. Three types of cross-features are constructed by combining path strength: industry-professional association features (such as...). ), skill requirement characteristics (such as , indicating skills (urgent need), characteristics of school quality (such as) (Combined professional employment rate and skills training coverage). Through mutual information entropy Features with strong relevance to demand forecasting targets are selected, and the top 70% of features with the highest mutual information values ​​are retained to reduce dimensionality redundancy. Knowledge-weighted interval standardization is employed, and key industry-related features and in-demand skill features are standardized using... Intervals, other features are used The interval, the standardized formula is ,in, , These are the upper and lower limits of the interval. The final output is well-organized data. ,in, 12 represents the time step, corresponding to 12 months; 68 represents the feature dimension, thus providing data for the prediction model.

[0052] S3: LSTM-Reinforcement Learning Fusion Prediction

[0053] This step will organize the data. The input fusion model achieves accurate prediction of vocational education needs through time-series modeling and dynamic optimization. The model input layer concatenates data features with entity embedding vectors; the entity embeddings are generated using the TransE algorithm. ,in, The input vector is composed of a head entity-relation-tail entity triple, ensuring that it contains both numerical features and semantic information. For example, the embedding vector of "Artificial Intelligence Application Major" is concatenated with the major's enrollment data and employment rate features to enhance the model's understanding of domain knowledge. The total dimension of the input vector is 68 (data features) + 32 (entity embedding dimension) = 100, thus ensuring that the model can simultaneously capture the numerical patterns and semantic relationships in the data.

[0054] Next, the dependencies between time-series data are captured through bidirectional propagation. The forward LSTM starts from time step... to Forward propagation focuses on learning trend features from historical data, such as enrollment trends in AI application majors over the past 12 months and changes in recruitment volume for corresponding positions. The hidden layer uses the Xavier initialization method and outputs the forward hidden state. Backward LSTM from time step to Backpropagation utilizes forward-looking information such as industry policy documents and vocational school development plans to output the backward hidden state. ; splice the two together (With 256 hidden units), the system comprehensively captures the dependencies between time-series data through bidirectional propagation, avoiding information loss caused by unidirectional propagation. The LSTM unit includes three gating mechanisms: input gate, forget gate, and output gate. The sigmoid activation function controls the inflow and outflow of information, effectively alleviating the gradient vanishing problem in long sequence training. For example, the forget gate can selectively forget short-term fluctuations that are irrelevant in the early stages (such as a temporary surge in enrollment caused by a policy pilot program).

[0055] Then, the influence of key elements on prediction is enhanced through dynamic weight allocation. The cosine similarity between the hidden state and the entity embedding vector at each time step is calculated. Measure time step Features and entities The degree of correlation; assigning high weights to key entities (such as emerging industries and skills in short supply) during the forecast period. and through Calculate attention weights to make the model pay more attention to time-step features associated with key entities (such as job demand data for the month the policy was released); the weighted sum is used to obtain the attention-enhanced hidden states. This highlights the contribution of key element correlation features to the prediction results, thereby enhancing the model's sensitivity to important information.

[0056] Furthermore, the adaptability of predictions is improved by dynamically adjusting model parameters. The agent is constructed using the Deep Deterministic Policy Gradient (DDPG) algorithm, and the state... Includes preceding prediction bias Deviation trend With data update volume (The amount of new job demand data in the current month) comprehensively reflects the model's predictive status and data changes; actions Adjustment amount for the number of hidden cells in LSTM Learning rate adjustment Attention weight coefficient adjustment amount Fine-tuning of parameters is achieved through a continuous action space; the reward function is designed as follows: This approach both penalizes prediction bias to improve accuracy and suppresses drastic parameter fluctuations to ensure stability. The agent learns the optimal adjustment strategy through an Actor-Critic network, which outputs actions. Critic network evaluates state-action value. The experience replay pool size was set to 10,000, and the target network soft update decay coefficient was set to 0.001 to ensure that the model can maintain stable predictive performance when data distribution changes (such as sudden changes in demand caused by industrial policy adjustments).

[0057] The multi-task output layer simultaneously predicts the quantity and structural characteristics of demand through parallel branches. The quantity prediction branch consists of a two-layer fully connected network (512 hidden layers, ReLU activation function), outputting the total vocational education demand for the next six months. (e.g., total enrollment demand for each month); the structural prediction branch consists of one convolutional layer (3×3 kernel size) and two fully connected layers, outputting the demand percentage for each industry and skill. (6 months, 15 industries, 5 skill levels). A multi-task loss function is employed. ,in, Mean square error (optimizing quantity prediction accuracy). Cross-entropy loss is used (to optimize the distribution matching of structure proportions). Prediction results are calibrated based on knowledge graph rules, and predictions exceeding a threshold are processed accordingly. Make corrections ( For industry (Output value share), ensuring that the forecast results conform to the objective laws of industrial development, and finally outputting the forecast sequence. .

[0058] S4: Generate a dynamic adjustment plan

[0059] This step is based on the prediction results. By combining knowledge constraints and optimization algorithms, optimal resource adjustment schemes are generated, transforming knowledge correlations into quantitative adjustment strategies. The calculation of demand changes is improved in accuracy by introducing industry correlation corrections; the basic change includes month-on-month changes. Difference from the actual value of the previous month), year-on-year change (Difference from the actual value in the same period last year), trend change (Reflecting the acceleration of change), it characterizes the dynamic features of demand from three dimensions. This is achieved through industry correlations within a knowledge graph. (Taking the industry and...) The mean correlation strength) is corrected for the change, and the result is obtained. For example, when the automotive industry ( When demand increases, it will be through correlation. Drive the automotive parts industry ( The increase in demand changes can be used to capture the impact of industrial linkages on the demand for vocational education, making the changes more in line with actual economic laws.

[0060] The dynamic coefficient matrix achieves differentiated resource allocation by integrating entity importance. (Admissions coefficient matrix) It is a diagonal matrix, with diagonal elements ,in, For industry In the The monthly output value as a percentage, and meet the requirements. , For industry Employment rate of the corresponding major The importance of industries is weighted at 1.2 for key industries and 1.0 for general industries. This ensures that enrollment adjustments favor high-value, high-employment, and high-importance industry-related majors. For example, the enrollment coefficient for majors related to the high-end equipment manufacturing industry is significantly higher than that for traditional industry majors. (Course coefficient) Defined as ,in, For skills The urgency of demand (the higher the growth rate, the stronger the urgency). For existing courses and skills The matching degree (calculated by the cosine similarity between the course outline and the skill description, with a value of 0-1). The importance weight of skills is assigned (1.3 for skills in short supply) to prioritize course adjustments that meet the urgent needs of skills with low existing matching rates. For example, when the demand for industrial robot operation skills surges and existing courses are insufficient to match the needs, the adjustment coefficient for the corresponding courses will increase significantly. Faculty coefficient. The calculation formula is ,in, For the current industry With skills The corresponding teacher-student ratio (the higher the teacher-student ratio, the larger the coefficient). For teachers' skills Proficiency level (assessed through training records and assessment results, with a value of 0-1). The importance weighting of industry and skills will be adjusted to focus teacher recruitment efforts on areas with high student-to-teacher ratios, low teacher proficiency, and high importance of teacher combination.

[0061] Constrained optimization uses mathematical modeling to find the optimal adjustment of resources, with the objective function set as maximizing resource utilization. ,in, The resource efficiency coefficient is calculated by combining the industry contribution rate and the graduate employment rate. Sectors with high industry contribution rates and high employment rates have larger efficiency coefficients, ensuring that resources are tilted towards high-value sectors. Constraints include three categories: upper limit constraints on enrollment quotas. ,in, The monthly enrollment limit approved by the education department. To ensure the current enrollment plan is not adjusted beyond the school's capacity limits, and considering the 3-month preparation time required for course development, the adjustment amount for newly added courses will be determined within the current month. Teacher staffing quota constraints ,in, The upper limit for the number of instructors corresponding to the industry and skills combination. The current number of teachers must comply with relevant personnel management regulations.

[0062] The Lagrange multiplier method is used to incorporate constraints into the objective function to construct the Lagrange function. ,in, These are Lagrange multipliers. (The last part is incomplete and likely refers to a different context.) , , Find the partial derivative and set it to 0. Then, using the KKT conditions (complementary relaxation), solve for the optimal adjustment amount: when hour, (Enrollment adjustments have reached the planned maximum); when hour, The resource adjustment amount is determined by the optimal solution of the objective function. Each plan corresponds to an adjustment scheme for enrollment plans, curriculum design, and faculty allocation, ensuring that the scheme maximizes resource utilization within the feasible range.

[0063] S5: Execution Feedback and Optimization

[0064] This step will adjust the amount. The system is transformed into executable tasks, and through closed-loop feedback, it continuously optimizes to achieve dynamic matching between vocational education resources and needs. Task decomposition is based on the relationships between subjects and tasks in a knowledge graph, breaking down adjustments into specific executable tasks. An example is the task of adjusting enrollment in a mechanical manufacturing major (…). =+50 people), based on the association in the knowledge graph that "the Admissions Office is responsible for the execution of professional enrollment plans", the task was assigned to the Admissions Office; the execution path was clarified, including conducting joint presentations with 3 equipment manufacturing companies (to be launched within 10 days of the release of the adjusted plan, with each session covering 200+ potential candidates) and placing targeted enrollment advertisements (budget of 20,000 yuan, a period of 2 weeks, covering 5 vocational high schools in the province); quantitative acceptance standards were set: the number of people reached by the presentations should not be less than 600, the enrollment completion rate should not be less than 95%, and the application review period should not exceed 7 working days. The task was pushed to the responsible person through the school's office management system. The system supports real-time progress reporting (such as daily updates of the number of applicants and uploading of company presentation confirmation letters) and feedback on questions. The department head responded to the execution difficulties submitted by the responsible person (such as conflicts in the presentation time of the cooperating companies) within 24 hours, and coordinated to adjust the execution path (such as changing the presentation companies or adjusting the time) to ensure the smooth implementation of the task.

[0065] Next, real-time monitoring of the execution process is achieved by integrating multi-source data and knowledge verification. A multi-dimensional execution data collection network is constructed: the campus card system collects the actual number of people reached by enrollment promotions (such as student QR code check-in data); the training equipment management system records the arrival progress of equipment required for course adjustments (such as the procurement, installation, and debugging status of PLC training equipment); and the teacher assessment system tracks the participation rate and assessment results of teacher training (such as the check-in rate and pass rate of industrial robot operation skills training). The real-time collected data is compared with the execution standard entities in the knowledge graph to calculate the deviation rate. When the deviation rate > When this happens, an early warning mechanism is triggered, and targeted corrective suggestions are pushed based on the knowledge graph.

[0066] Furthermore, the system is dynamically optimized through continuous iteration. The execution deviation rate will be... Input the reinforcement learning agent S3 and update its state. And adjust the movements If the prediction bias consistently exceeds the threshold, increase the number of LSTM hidden units to improve the model's fitting ability; if the course adjustment bias is large, increase the course coefficient weight of the corresponding skill. This strengthens the responsiveness of curriculum adjustments to changing demands. Simultaneously, the knowledge graph is iteratively updated based on the causes of deviations: if the prediction deviation for a certain skill demand remains significant, the correlation between that skill and related industries is strengthened. If there are significant deviations in enrollment execution, adjust the correlation between promotional channels and majors, reduce the weight of ineffective online advertising channels, and increase the weight of corporate presentation channels. Conduct a full-process optimization evaluation every quarter, and verify the optimization effect by comparing the predicted MAE (mean absolute error) and resource matching rate (the proportion of resources that meet the needs after adjustment) before and after optimization.

[0067] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A vocational education demand forecasting system based on LSTM time series forecasting, characterized in that, It includes a multi-source data acquisition module, a knowledge graph construction module, an intelligent data processing module, an LSTM-reinforcement learning prediction module, a dynamic adjustment module, and an execution feedback module; The multi-source data acquisition module connects to government industry databases, recruitment information platforms, and vocational school management systems through customized interfaces to collect three types of data: industry economic indicators, job demand information, and education and school operation data. The knowledge graph construction module uses natural language processing technology to extract core entities from text data, combines rules in the field of vocational education to build entity relationships, calculates the association strength through entity co-occurrence frequency, and binds numerical data as attributes to corresponding entities to form a knowledge system containing semantic associations and quantitative information. The intelligent data processing module combines knowledge graph association rules to identify and correct time-series outliers, fills data gaps with strong association entity features, and generates high-order features based on entity association paths. The LSTM-reinforcement learning prediction module uses a bidirectional LSTM network to capture the temporal dependence of data, introduces a knowledge attention mechanism to strengthen the feature weights of key elements, and dynamically optimizes the model parameters by combining reinforcement learning algorithms. It outputs the prediction results of the quantity and structure of demand through multi-task learning and calibration. The dynamic adjustment module constructs a dynamic coefficient matrix for enrollment plans, curriculum design, and faculty allocation based on the importance weights of entities in the knowledge graph. It transforms industry linkage constraints and school resource limits into mathematical constraints and solves the adjustment amount that maximizes resource utilization through optimization algorithms. The execution feedback module decomposes and adjusts the tasks into specific tasks based on the relationship between the responsible entities and tasks in the knowledge graph and clarifies the execution standards. It monitors the execution data in real time, calculates the deviation rate, and feeds the deviation information back to the front-end module to optimize the model parameters and knowledge association rules.

2. The vocational education demand forecasting system based on LSTM time series forecasting as described in claim 1, characterized in that, The multi-source data acquisition module collects industrial economic indicators including the monthly output growth rate of key industries, the proportion of industrial structure, and the regional unemployment rate. These indicators are obtained periodically through the government open platform API interface, with the collection cycle dynamically set according to the data granularity. Job demand information covers the monthly recruitment volume, skill set, and average salary for occupational positions. It uses distributed web crawlers to collect data from mainstream recruitment platforms and filters out invalid positions without clear skill requirements, and labels the industry to which the positions belong. Educational data includes the number of students enrolled in each major, the graduate employment rate, and the core curriculum settings. The data is obtained in an anonymized manner, retaining only the summary data at the major level.

3. The vocational education demand forecasting system based on LSTM time series forecasting as described in claim 1, characterized in that, The knowledge graph construction module uses a BERT-based pre-trained model combined with an occupational classification dictionary to extract four types of entities: industry, job, major, and skill. Based on entity co-occurrence frequency and domain rules, it constructs three types of association relationships: industry includes job, job requires skills, and major training skills. The association strength is calculated and normalized to the 0-1 range. Numerical data such as job recruitment volume, salary, major enrollment, and employment rate are bound to the corresponding job entity and major entity, respectively.

4. The vocational education demand forecasting system based on LSTM time series forecasting as described in claim 1, characterized in that, When identifying time-series outliers, the intelligent data processing module uses a local weighted regression algorithm to fit the trend curve, constructs a dynamic threshold based on industry volatility, and verifies the consistency between the data and the trends of related entities through a knowledge graph. If there is a discrepancy, it uses the average value of the same industry and similar positions for correction. When filling data gaps, random gaps are filled by a method that combines time-series interpolation with the growth rate of strongly related entities, while systematic gaps are filled by a two-layer GraphSAGE algorithm to aggregate and predict the features of related entities. When generating higher-order features, the module calculates the path strength based on the entity association path, constructs industry-professional association features, skill demand features, and school quality features, filters features through mutual information entropy, and uses knowledge-weighted interval standardization.

5. A dynamic adjustment method for a vocational education demand forecasting system based on LSTM time series forecasting as described in any one of claims 1 to 4, characterized in that, Includes the following steps: S1: Multi-source data collection and knowledge graph construction, cross-channel aggregation of industrial economic data, job demand data, and vocational education data, extraction of core entities and construction of relationships, forming a system that integrates knowledge and data; S2: Intelligent data processing, based on a fusion system, identifies and corrects outliers, fills in missing data, generates and filters high-order features, and outputs well-organized data; S3: LSTM-Reinforcement Learning Fusion Prediction inputs regularized data into the fusion model, concatenates data features and entity embedding vectors, captures temporal dependencies through bidirectional LSTM, introduces knowledge attention mechanism and reinforcement learning to optimize the model, and outputs calibrated demand prediction results. S4: Generate a dynamic adjustment plan, calculate the change in demand based on the forecast results and correct it through the industrial correlation, construct a dynamic coefficient matrix, and solve the optimal resource adjustment amount in combination with the constraints. S5: Execution feedback and optimization, decompose the adjustment amount into specific tasks and execute them, monitor the execution process in real time to calculate the deviation rate, and provide feedback on deviation information to optimize the model and knowledge graph.

6. The method for dynamically adjusting vocational education demand based on LSTM time series prediction as described in claim 5, characterized in that, In step S1, during data collection, monthly data is collected on a fixed date each month, and quarterly data is collected on a fixed date in the first month, with the data statistical range and time stamp recorded synchronously. During knowledge graph construction, the entity association strength calculation formula is normalized to a value range of 0-1, integrating knowledge and data into a system. ,in, For knowledge graphs, , , These are data matrices for industrial economics, job demand, and vocational education.

7. The method for dynamically adjusting vocational education demand based on LSTM time series prediction as described in claim 5, characterized in that, In step S2, a fixed smoothing parameter is used when fitting the trend curve, and a Gaussian kernel function with a specific bandwidth is selected for weight calculation to balance trend capture and short-term fluctuations. The dynamic threshold is determined based on the volatility of the industry's output value over the past 6 months. When filling random missing data, the month-on-month growth rate is first calculated using data from the preceding and following periods to obtain a preliminary estimate. Then, the growth rate of entities with sufficient correlation strength is combined to obtain a corrected value. The final filling value is generated by superimposing the weights. When imputing systematic missing data, entities are treated as nodes, historical data and related entity attributes are used as features, and the association strength is used as the edge weight. The predicted imputation value is output by aggregating neighbor node information and inputting it into a fully connected layer. In the feature standardization stage, different numerical ranges are defined for key industry-related features, shortage skill features, and general features. Specifically, key industry-related features and shortage skill features are subject to different numerical ranges. Intervals, other features are used The interval is calculated, and finally, regularized data is output.

8. The method for dynamically adjusting vocational education demand based on LSTM time series prediction as described in claim 5, characterized in that, In step S3, entity embeddings are generated using the TransE algorithm, and the input vector dimension is the sum of the data feature dimension and the entity embedding dimension. In the bidirectional LSTM, the forward LSTM learns historical trend features by forward propagation from the start time step to the end time step, and the hidden layer is initialized using Xavier. The backward LSTM integrates look-ahead information by backward propagation from the end time step to the start time step. The hidden states output by both are concatenated into a vector with a fixed number of hidden units. The knowledge attention mechanism assigns high weights to key entities and calculates attention weights by calculating the cosine similarity between the hidden state and the entity embedding vector, thus obtaining attention-enhanced hidden states. Reinforcement learning employs the DDPG algorithm to construct the agent. The state includes the preceding prediction bias, bias trend, and data update amount. Actions include adjustments to the number of LSTM hidden units, learning rate, and attention weight coefficients. The reward function simultaneously penalizes prediction bias and drastic parameter changes. In the multi-task output layer, the quantity prediction branch is a 2-layer fully connected network, and the structure prediction branch is a combination of a 1-layer convolutional layer and a 2-layer fully connected layer. The multi-task loss function is optimized, and predictions exceeding the threshold are corrected based on knowledge graph rules.

9. The method for dynamically adjusting vocational education demand based on LSTM time series prediction as described in claim 5, characterized in that, In step S4, the change in demand includes month-on-month change, year-on-year change, and trend change, which is corrected by the average industrial correlation of the past 12 months in the knowledge graph; in the dynamic coefficient matrix, the enrollment coefficient matrix is ​​a diagonal matrix, and the diagonal elements are calculated by combining the proportion of industrial output value, professional employment rate and industrial importance weight, with the weight of key industries being higher than that of general industries. The course coefficient is calculated by combining the urgency of skill demand, the matching degree between existing courses and skills, and the importance of skills, with higher weighting for skills in short supply; The teacher coefficient is calculated by combining the teacher-student ratio, teacher skill proficiency, and the weights of industry and skill importance. The objective function of the constraint optimization is to maximize resource utilization. The constraints include the upper limit of enrollment plan, course adjustment cycle, and teacher staffing limit. The constraints are incorporated into the objective function using the Lagrange multiplier method, and the optimal adjustment amount is solved by combining the KKT conditions.

10. The method for dynamically adjusting vocational education demand based on LSTM time series prediction as described in claim 5, characterized in that, In step S5, the execution path and quantitative acceptance criteria are clearly defined during task decomposition. Tasks are pushed through the school's office management system, which supports progress reporting and feedback on questions. Execution difficulties are addressed within a specified time. Execution monitoring collects multi-dimensional data through the campus card system, training equipment management system, and teacher assessment system. The deviation rate is calculated by comparing the data with the execution standard entities in the knowledge graph. When the deviation rate exceeds a threshold, an early warning is triggered and corrective suggestions are pushed. During system optimization, the deviation rate is input into the reinforcement learning agent to adjust model parameters. The association strength of the knowledge graph is updated based on the cause of the deviation. The predicted MAE and resource matching rate before and after optimization are compared quarterly to verify the optimization effect.

Images