Construction method and system of industrial production prediction model

Abstract

The invention relates to the technical field of Internet of Things, and provides an industrial production prediction model construction method and system. A standard industrial data set is subjected to periodic decomposition and trend smoothing processing according to a time period to generate a time sequence change sequence, and the time sequence change sequence is subjected to power load segmentation adjustment and feature extraction to obtain a production power load capacity sequence and a capacity feature set. Carrying out industrial state type identification and state probability calculation according to the production power load sequence and the capacity feature set to generate a state probability sequence, and integrating the production power load sequence and the capacity feature set according to the state probability sequence to obtain a prediction variable set of each industrial state type; and constructing an industrial production prediction model according to the state probability sequence and the prediction variable set. According to the method, the state condition regression model is constructed through temperature and calendar stripping, mixing fusion and state recognition, and the precision, robustness and interpretability of industrial output proximity prediction are improved.

Description

Technical Field

[0001] This application relates to the field of Internet of Things (IoT) technology, and in particular to a method and system for constructing an industrial production prediction model. Background Technology

[0002] The requirements and importance of industrial production forecasting lie in providing high-precision short-term and nowcast forecasts to support production scheduling, energy management, and supply chain coordination, while also providing uncertainty metrics and causal interpretability to enable decision-makers to weigh risks and take intervention measures. Therefore, forecasting systems must ensure data integrity and time alignment, integrate high-frequency and low-frequency signals, isolate seasonal and calendar interference, and consider environmental drivers. They must also possess online update capabilities and rapid response to sudden structural changes. The results directly impact capacity planning, inventory control, and emergency management, affecting enterprise operational efficiency and cost control.

[0003] Existing industrial production forecasting models are constructed in various ways, typically including classic statistical models based on time series (such as ARIMA and its seasonally adjusted variants), causal models based on regression (using meteorological, calendar, and economic indicators as explanatory variables), mixed-frequency regression or simple leading indicator regression, and black-box methods based on machine learning (such as random forests, gradient boosting, and neural networks) to directly map input features to output indicators. In engineering practice, it is also common to perform simple smoothing or feature summarization on power or production line signals and use them as exogenous inputs to improve short-term performance. However, these methods differ significantly in handling multi-frequency data fusion, state dependencies, and uncertainty quantification.

[0004] Most existing methods either crudely exclude weekends / holidays, resulting in information loss, or simply treat seasonality as a static control variable, failing to capture the non-stationarity of coefficients under sudden economic or production changes. They also lack the ability to dynamically integrate leading signals such as high-frequency power with state-aware mixing weights and synchronously quantify and predict uncertainties. As a result, they are significantly inaccurate during sudden disturbances or structural transitions and cannot provide reliable risk warnings. Summary of the Invention

[0005] In view of this, this application provides a method and system for constructing an industrial production prediction model to solve the problem of insufficient ability to indicate sudden change risks due to inadequate comprehensive processing of calendar effects, high-frequency signals and system state switching.

[0006] The first aspect of this application provides a method for constructing an industrial production forecasting model, the method comprising: The received industrial dataset is preprocessed to obtain a standard industrial dataset, which is then periodically decomposed and trend smoothed according to a preset time period to generate a time-series change sequence. The power load is segmented and its features are extracted from the time-series variation sequence to obtain the power load sequence and capacity feature set; Based on the power load sequence and the capacity feature set, the industry status type is identified and the status probability is calculated to generate a status probability sequence. Based on the state probability sequence, the production power load sequence and the capacity feature set are integrated to obtain a set of predictive variables for each industry state type; An industrial production prediction model is constructed based on the state probability sequence and the set of prediction variables.

[0007] In an optional implementation, the preprocessing of the received industrial dataset to obtain a standard industrial dataset includes: The received industrial dataset is time-domain aligned to obtain a time-synchronized dataset; Based on the time sequence, the time-series synchronized dataset is subjected to abnormal data identification and cleaning, as well as interpolation compensation for missing data, to obtain a clean industrial dataset. The importance of the clean industrial dataset is assessed based on the preset production planning data to obtain the importance labels corresponding to each time point. The weight coefficients corresponding to each time point are obtained from the preset set of weight coefficients based on the importance markers; The data in the clean industrial dataset are weighted and corrected according to the weighting coefficients to obtain a standard industrial dataset.

[0008] In an optional implementation, the step of performing periodic decomposition and trend smoothing on the standard industrial dataset according to a preset time period to generate a time-series change sequence includes: The production capacity-related dataset is periodically decomposed according to a preset time period using a preset seasonal adjustment algorithm to obtain industrial production capacity data sequences for each time period. The industrial capacity data sequence is processed by a preset differential algorithm to generate production trend data sequences for each time period. Heating and cooling demands are calculated based on preset base temperature values ​​and the ambient temperature dataset to obtain temperature regulation indices for each time period. The temperature regulation index and the production trend data sequence are aligned and merged according to the time sequence to generate a time-series change sequence.

[0009] In an optional implementation, the step of segmenting and extracting features from the time-series variation sequence of power load to obtain the production power load sequence and the capacity feature set includes: The non-production power load adjustment amount is calculated in each time period based on the temperature regulation index using a preset temperature and load correlation algorithm. Based on the power load adjustment amount, the production trend data sequence is segmented and corrected for power load to obtain the production power load sequence. The time-series change sequence is processed to extract industrial output change features to obtain a capacity feature set.

[0010] In an optional implementation, the step of identifying industry state types and calculating state probabilities based on the production power load sequence and the capacity feature set, and generating a state probability sequence, includes: Step S31: Based on the preset time point set, group the production power load sequence and the production capacity feature set to obtain the state identification sample group for each time point. Step S32: Randomly label the state identification sample group with preset industry state types using a preset random allocation method to generate an industry state type sequence corresponding to each time point; Step S33: Perform time-series change statistical analysis on the state identification sample group of the same industry state type to obtain the capacity and load correlation index corresponding to each industry state type, and perform industry state type transition frequency statistics on the industry state type sequence according to the time sequence to obtain the transition probability between states. Step S34: Using a preset likelihood function algorithm, calculate the expected probability of each time point belonging to each industry state type based on the capacity and load correlation index and the state transition probability. Step S35: Update the capacity and load correlation index and state transition probability of the state identification sample group at the corresponding time point according to the expected probability using a preset maximization algorithm. Step S36: Calculate the first change in the capacity and load correlation index before and after the update, and calculate the second change in the state transition probability before and after the update; Repeat steps S34 to S36 until both the first change and the second change are less than the preset change threshold, and use the expected probability as the state probability at the corresponding time point. Step S37: Determine the industry state type at the corresponding time point based on the preset probability threshold and the state probability; Step S38: Arrange and combine the state probabilities and the industry state types according to the time sequence to generate a state probability sequence.

[0011] In an optional implementation, the step of integrating the production power load sequence and the capacity feature set based on the state probability sequence to obtain a set of predictive variables for each industry state type includes: The state probability is used as the fusion weighting coefficient of the corresponding load in the power production load sequence; Based on the preset frequency period and the fusion weighting coefficient, the load quantities in the power production load sequence are weighted and fused to obtain a composite power load sequence. Based on the frequency period and the industry status type, the composite power load sequence, the status probability sequence, and the capacity feature set are grouped and integrated to obtain a set of predictive variables for each industry status type.

[0012] In an optional implementation, constructing an industrial production prediction model based on the state probability sequence and the set of predictor variables includes: The set of predictive variables corresponding to the industry status type is used as the prediction sample dataset for each industry status type. By using a preset steady-state estimation regression algorithm, regression analysis is performed on the predicted sample dataset to obtain variable prediction correlation data for each industry state type; Based on the predicted correlation data of the variables and the predicted sample dataset, regression equation models for each industry status type are constructed to obtain industrial production prediction models.

[0013] A second aspect of this application provides an apparatus for constructing an industrial production prediction model, the apparatus comprising: The data optimization module is used to preprocess the received industrial dataset to obtain a standard industrial dataset. The decomposition and smoothing module is used to perform periodic decomposition and trend smoothing on the standard industrial dataset according to a preset time period to generate a time-series change sequence. The feature extraction module is used to perform segmented adjustment and feature extraction on the time-series change sequence of power load to obtain the power load sequence and capacity feature set; The state analysis module is used to identify the industry state type and calculate the state probability based on the production power load sequence and the capacity feature set, and generate a state probability sequence. The variable integration module is used to integrate the production power load sequence and the capacity feature set according to the state probability sequence to obtain a set of predictive variables for each industry state type. The model building module is used to build an industrial production prediction model based on the state probability sequence and the set of prediction variables.

[0014] A third aspect of this application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for constructing an industrial production prediction model as described above.

[0015] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for constructing an industrial production prediction model as described above.

[0016] In summary, this application includes at least the following beneficial technical effects: 1. By performing complete data preprocessing, separating working day weights from non-productive energy consumption, and combining high-frequency power signals with frequency mixing, the model can more accurately map recent high-frequency fluctuations to monthly output, thereby improving the accuracy and timeliness of near-term and short-term forecasts.

[0017] 2. By introducing Markov state transition and state condition equations, the model can identify and distinguish between two operating modes: "low volatility" and "high volatility," and adjust the coefficients and confidence levels accordingly, thus maintaining good performance during periods of catastrophe.

[0018] 3. By stripping away temperature and calendar effects and outputting state probabilities and uncertainty decomposition, the model not only provides point predictions but also state probabilities, confidence intervals, and sources of uncertainty, making the prediction results more interpretable and facilitating targeted operation and maintenance or policy decisions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating a method for constructing an industrial production prediction model provided in an embodiment of this application; Figure 2 This is a functional block diagram of an apparatus for constructing an industrial production prediction model provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] like Figure 1 The diagram shows a flowchart of a method for constructing an industrial production forecasting model according to an embodiment of this application. The method for constructing an industrial production forecasting model according to an embodiment of this application includes the following steps.

[0023] Step S1: Preprocess the received industrial dataset to obtain a standard industrial dataset, and perform periodic decomposition and trend smoothing on the standard industrial dataset according to a preset time period to generate a time-series change sequence.

[0024] It should be understood that industrial datasets are not simple records from a single source, but rather a collection of raw time-series data closely related to industrial production activities, systematically collected and aggregated from multiple authoritative data sources. Specifically, this collection includes: electricity load sequences consisting of hourly data on total electricity consumption of regional industrial users or key enterprise electricity consumption by detail, obtained from the State Grid Corporation's open data platform or regional power trading centers; sequences of key meteorological variables such as temperature, humidity, and precipitation, obtained from the National Meteorological Administration or professional service agencies, covering major meteorological stations in the target industrial area; sequences of economic indicators such as industrial production indices, capacity utilization rates, and order indices, reflecting the level of macro-industrial activity, extracted from databases of the National Bureau of Statistics, industry associations, or international organizations; sequences of calendar information, labeled with daily working status (such as normal working days, weekends, statutory holidays, and adjusted working days) and seasonal production characteristics (such as peak seasons and off-seasons), parsed from statutory holiday announcements, industry production cycle arrangements, and historical work calendars; and sequences of industry parameters for specific industrial sectors (such as blast furnace operating rates in the steel industry and raw material conversion rates in the chemical industry), collected from industry monitoring reports and enterprise ERP or MES systems. The aforementioned multi-source heterogeneous data set forms the original data foundation for building the predictive model. Its comprehensiveness and complementarity aim to capture the multi-dimensional factors affecting industrial production in all aspects.

[0025] Because data from different sources may have different timestamp formats, sampling frequencies (such as hourly, daily, and monthly), or time zone differences, direct merging will lead to time misalignment and analytical distortion. Therefore, after receiving industrial datasets, time-domain alignment is the first step. This application establishes a unified and accurate timeline, transforming and unifying all data to a standard timestamp based on "calendar day," and creating a multi-level time index covering days, weeks, months, quarters, and years. This ensures that a record of electricity load data on the same date (e.g., July 15, 2023) can accurately correspond to information such as the average temperature of that day, the industrial output index of that month, and whether that day is a workday with adjusted working hours, thereby obtaining a time-series synchronized dataset that is completely synchronized in the time dimension.

[0026] The raw data may contain outliers due to equipment failure, communication errors, or human input mistakes during collection or transmission, and may also have missing data at certain points in time. Without proper processing, these data quality issues will severely interfere with the model's identification of real industrial activity patterns. After obtaining the time-series synchronized dataset, this application immediately performs outlier identification and cleaning, as well as interpolation compensation for missing data, to obtain a clean industrial dataset. In this application, outlier identification and cleaning typically employs a strategy combining statistical methods and domain knowledge. For example, for power load sequences, the 3σ principle can be applied to initially identify values ​​deviating from the norm, and then verified using historical event records (such as known equipment maintenance shutdowns or power rationing caused by extreme weather). Outliers confirmed as erroneous data are replaced with the weighted average of data from adjacent time periods. For interpolation compensation of missing data, the degree of missing data needs to be assessed. When the proportion of missing values ​​in a single sequence is small (e.g., no more than 3% of the total data points) and the missing pattern is random, linear interpolation can be used to smooth the data based on known data values ​​before and after the missing points to maintain the continuity of the sequence. For example, if the temperature data for a certain day is lost due to sensor failure, a reasonable estimate for that day can be filled in by linear calculation based on the measured temperatures of the previous and following days.

[0027] Subsequently, to quantify the actual contribution of each calendar day to industrial production activities, this application uses production planning data to assess the importance of the data corresponding to each calendar day in the clean industrial dataset. The production planning data reflects calendar information on the actual production intensity on different dates, particularly the workday arrangement. The data assessment operation used in this application assigns an importance label to each time point, which is directly related to the work status attribute of that day. For example, for a data record labeled "legal working day," its initial importance label may indicate that it is a standard production day; while for a record labeled "legal holiday" but belonging to a continuous production industry such as chemical industry, its importance label needs to reflect the characteristic that the industry may still maintain high production capacity during the holiday. Next, based on the generated importance labels, the system dynamically obtains the specific weight coefficients corresponding to each time point from a preset set of weight coefficients. The set of weight coefficients is a predefined set of rules that maps different work status types to specific numerical coefficients, and its numerical settings comprehensively consider legal arrangements, industry practices, and historical data patterns. For example, the weighting coefficient for statutory working days is often set at a baseline value of 1.0; the weighting for statutory weekends in the manufacturing industry may be set at 0.3 to reflect its characteristics of shutdown or reduced production on weekends; while the weighting for weekends in process industries (such as chemicals) may be as high as 0.8 to reflect its continuous production mode; the weighting for statutory holidays is usually set between 0.1 and 0.4, with the specific value adjusted according to the importance of the holiday and the industry's historical operating rate.

[0028] It should be understood that the uneven distribution of the calendar structure of heterogeneous data over time causes systematic and non-productive fluctuations in industrial production and related economic and power data sequences. The data distribution structure disturbances present in the clean industrial dataset are called calendar effects. This effect does not reflect changes in actual production efficiency or market demand, but rather stems from data distortion caused by fixed social schedules and administrative arrangements, disrupting the direct comparability between data from different time periods (such as different months or quarters). Without correction, the model will misinterpret such disturbances as real changes in industrial activity, leading to distorted predictions. Therefore, after obtaining the weight coefficients corresponding to each data point in the clean industrial dataset, a weighted correction is performed on the data in the clean industrial dataset to eliminate the calendar effect between key data points. The weighted correction operation in this application is particularly aimed at indicators such as power load and industrial output, which are directly affected by the number of production days. Its core is to eliminate fluctuations caused by differences in calendar structure such as different month lengths and uneven distribution of holidays. The correction method does not simply remove non-working day data, but introduces the concept of "working day equivalent" for calculation. For example, the raw total electricity load for a given month is not used directly. Instead, it is divided by the sum of the weighting coefficients for all days in that month (i.e., the "working day equivalents" for that month) to obtain the "working day equivalent load." This operation makes electricity load data between different months directly comparable because the data has been standardized to the same "effective workload" benchmark. The output after this weighted correction process is the standard industrial dataset that has undergone calendar effect correction.

[0029] It should be understood that the standard industrial dataset obtained above can be logically divided into a capacity-related dataset for characterizing production activities (mainly including the corrected industrial output index, electricity load, etc.) and an ambient temperature dataset for characterizing environmental factors. In order for the constructed model to focus on the capacity-related dataset, it is necessary to further extract the inherent time-series features of the standard industrial dataset.

[0030] First, this application uses a pre-defined seasonal adjustment algorithm to perform periodic decomposition on key sequences in the capacity-related dataset, such as industrial capacity data sequences, based on a pre-defined time period (usually a 12-month annual cycle). The seasonal adjustment algorithm (e.g., the X-13ARIMA-SEATS algorithm) is a statistical calculation method specifically designed to identify, estimate, and remove fixed periodic fluctuation components (i.e., seasonal factors) from time series data. This algorithm is used to purify the regular fluctuations that commonly occur in industrial production and related economic data, which repeat at fixed periods (e.g., annual, quarterly, monthly), thereby exposing the more essential trend-cyclical components and irregular disturbance components in the sequences. Industrial production activities inherently exhibit a series of periodic patterns driven by natural laws, social customs, or administrative arrangements. These patterns can cause predictable fluctuations in the capacity-related dataset that are unrelated to macroeconomic conditions or actual production efficiency. If these fluctuations are not separated, they will severely interfere with the model's identification of key signals such as economic inflection points and trend changes. For example, the output and electricity consumption of the air conditioning manufacturing industry typically peak before summer, which is related to the annual cycle of surging market demand in summer; similarly, during the Spring Festival, due to the nationwide long holiday, the output and electricity consumption of almost all industries plummet to their annual low. These recurring peaks and troughs within the year constitute strong seasonal signals. The seasonal adjustment algorithm used in this application first performs a preliminary analysis on the capacity-related dataset (e.g., unadjusted monthly industrial output index) to identify the length of its seasonal cycle based on a pre-set time period (e.g., data with an annual time period of 12 months). Next, a model containing trend, seasonal, and irregular components is constructed, and through repeated estimation and verification of the capacity-related dataset, it is precisely decomposed into trend-cyclical components, seasonal components, and irregular components (i.e., industrial capacity data sequences). For the capacity-related dataset processed in this application, which typically exhibits seasonal fluctuations proportional to the magnitude and level, the seasonal adjustment algorithm uses a multiplicative model to decompose it (i.e., the obtained industrial capacity data sequence = capacity-related dataset / seasonal component estimated by the algorithm). Taking the monthly crude steel production series in the steel industry as an example, capacity-related datasets may show production peaks from March to May and September to November each year. Seasonal adjustment algorithms can accurately estimate the specific values ​​of these cyclical peaks (i.e., the seasonal component) and subtract them from the capacity-related dataset. The adjusted industrial capacity data series may reveal a gradual, long-term downward trend after removing the regular spring and autumn production peaks, potentially reflecting actual industry overcapacity or a shift in demand structure. Without this seasonal adjustment, the constructed model might incorrectly interpret the annual spring production rebound as a signal of economic recovery.The aforementioned periodic decomposition operation removes non-trend, recurring seasonal fluctuations, thereby obtaining a seasonally adjusted industrial capacity data sequence, making the long-term trend and short-term abnormal changes in the sequence more prominent. However, the seasonally adjusted industrial capacity data sequence may still contain long-term trends, causing the sequence to fail to meet the stationarity requirements for statistical modeling. Therefore, this application also requires trend smoothing of the sequence using a pre-defined differencing algorithm. The differencing operation used in this application calculates the change in values ​​between adjacent time points, rather than the absolute level value. For example, performing a first-order differencing on the monthly industrial output index sequence, i.e., calculating the difference between the index value of this month and the previous month, yields a month-on-month output change sequence. Through further trend smoothing, the long-term trend of the sequence is effectively eliminated, shifting the focus from "output level" to "output change." The resulting sequence is called a production trend data sequence, which has more stable statistical characteristics and is more suitable for subsequent regression and state identification modeling.

[0031] In industrial production environments, maintaining a baseline comfort temperature (typically within the range of 18-22°C, with minor adjustments based on specific industries and regional climates) is a key engineering requirement for ensuring production safety, equipment stability, personnel efficiency, and consistent product quality. For example, the positioning accuracy of precision CNC machine tools is required to be at the micrometer level; a 1°C change in ambient temperature can lead to positioning deviations exceeding 10 micrometers due to thermal deformation of the machine tool structure. The rate constant of chemical reactions follows the Arrhenius equation, exhibiting an exponential relationship with temperature; fluctuations in reactor temperature directly affect product purity and yield. This requirement directly impacts the reliability and economy of the production system, necessitating the separate separation of temperature's influence on electrical load when constructing predictive models. Specifically, this application uses a pre-set baseline temperature value and a daily average temperature sequence from an ambient temperature dataset to calculate the energy demand for heating and cooling (i.e., the temperature regulation index) daily. When the daily average temperature is lower than the baseline temperature, the difference between the baseline temperature and the daily temperature is calculated as the "heating" temperature regulation index for that day; when the average temperature is higher than the baseline temperature, the difference between the daily temperature and the baseline temperature is calculated as the "cooling" temperature regulation index. These two indices quantify the energy requirements for heating and cooling to maintain a baseline comfort temperature, and can more accurately capture the impact of temperature on auxiliary power loads such as air conditioning and ventilation in industrial production.

[0032] Finally, based on a unified time sequence, the production trend data series representing changes in production activities were aligned and merged with the temperature regulation index representing environmental drivers. The alignment operation ensured that data within the same row were perfectly matched at any given time point; the merging operation integrated these key variables into the same data table, forming the final time-series change sequence used for subsequent modeling and analysis. This sequence combines stationarity with explanatory power, eliminating interference from calendar effects, seasonal cycles, and long-term trends while retaining core information on short-term changes in production activities and the nonlinear impact of environmental temperature. This lays a solid data foundation for the next step of accurately separating the effect of temperature on electricity load.

[0033] Step S2: Perform segmented adjustment and feature extraction on the time-series change sequence of power load to obtain the production power load sequence and capacity feature set.

[0034] Step S2 of this application requires further separating and extracting the core signals that purely reflect the intensity of industrial production activities from the pre-purified time-series variation sequence. The time-series variation sequence, as input, has undergone calendar effect correction, seasonal decomposition, and trend smoothing, and integrates a temperature regulation index characterizing the impact of ambient temperature. However, the electricity load data in this sequence still contains temperature control energy consumption generated to maintain a baseline comfortable temperature for the production environment. This energy consumption fluctuates with weather changes and has a weak correlation with actual production activities. If not separated, the model will misjudge the surge in electricity consumption caused by cold waves or heat waves as an expansion of production scale, leading to systematic bias in prediction. Therefore, the core objective of this step is to accurately quantify and eliminate this portion of "non-production-required electricity load" through a "temperature and load correlation algorithm," and simultaneously prepare a "capacity feature set" to characterize production results.

[0035] First, using a preset temperature-load correlation algorithm, the amount of non-production-related power load adjustment required in each time period is accurately calculated based on the pre-calculated temperature regulation index. In this embodiment, the temperature regulation index is defined as the heating degree-day index and the cooling degree-day index, which quantitatively represent the theoretical energy demand required to maintain the daily average temperature above or below a preset baseline temperature value (e.g., 18°C). It should be understood that the temperature-load correlation algorithm used in this application is a statistical modeling tool used to reveal the quantitative relationship between changes in power load and changes in the temperature regulation index. Since this relationship is usually non-linear—for example, in extremely cold weather, the increase in power load resulting from each degree increase in heating demand may differ from that in mild weather, and even when the temperature is within the comfort range, the temperature control load may approach zero—the temperature-load correlation algorithm used in this application employs piecewise linear regression to capture the differentiated impacts of different temperature ranges. Specifically, the algorithm constructs a regression equation, using the change in electricity load over a target time period (e.g., in weeks) as the dependent variable, and the changes in the heating and cooling daily temperature indices, as well as the weights of weekdays (which may serve as control variables), as independent variables. Through estimation methods such as ordinary least squares, regression coefficients reflecting the intensity of temperature influence can be obtained. For example, for a regional power grid with a concentration of manufacturing enterprises, the fitting results might show that for every unit increase in the cooling daily temperature index each week, the total grid load increases by an average of 5 megawatts; while for every unit increase in the heating daily temperature index, the load increases by an average of 8 megawatts. These coefficients are multiplied by the actual change in the temperature regulation index within the corresponding period, and the result is the estimated change in electricity load caused by temperature fluctuations, i.e., the "non-production-required electricity load adjustment." This adjustment is not a fixed value but a dynamically changing sequence over time, accurately depicting the "interference" pattern of weather factors on the electricity consumption curve.

[0036] After calculating the dynamic power load adjustment, the power load of the production trend data sequence, which represents the power consumption trend of production activities, needs to be segmented and corrected based on this adjustment to obtain a pure production power load sequence. It should be understood that the production trend data sequence obtained in step S1 reflects the short-term trend of the load itself after removing calendar and seasonal factors. The segmented correction operation used in this application involves subtracting the power load adjustment calculated for the corresponding time period from the production trend data sequence point by point; that is, production power load sequence = production trend data sequence - power load adjustment. By removing the portion of the load data that can be explained by temperature variables, the remaining portion is mainly attributed to changes in actual production activities, fluctuations in equipment efficiency, and other random factors not covered by temperature. Taking the automotive manufacturing industry as an example, a week in which high temperatures may cause a significant increase in electricity consumption for air conditioning in the factory, resulting in a peak in the original load sequence. Through the above corrections, the temperature control load adjustment calculated based on the extremely high daily cooling index of the week is subtracted from the total load change. The remaining production power load sequence shows a stable or even slightly declining trend, which may truly reflect the actual reduction in production pace due to production line maintenance during the week. The production power load sequence obtained through these segmented corrections is a cleaner and more direct indicator linked to core production behaviors such as industrial machine operation and assembly line operation, providing high-quality feature input for subsequent identification of economic status.

[0037] Simultaneously, step S2 of this application also requires extracting a capacity feature set from the time-series variation sequence. The feature extraction operation employed in this application focuses on indicators that directly represent industrial production results, such as the industrial output index sequence that has undergone seasonal adjustment and differencing. Specifically, key data is purposefully screened and formatted from the processed time-series variation sequence. Since the industrial output data in the time-series variation sequence has undergone the same stabilization preprocessing (seasonal adjustment and differencing) as the load data, they are feature sequences existing in the form of changes, such as the month-on-month growth rate of industrial output. The extraction process involves separating these sequences, and possibly other related stabilization economic indicators (such as changes in capacity utilization), from the time-series variation sequence and organizing them according to a unified time-series index to form a dedicated capacity feature set. The capacity feature set provides target variables and necessary control variables for model construction. For example, for a regional chemical industry forecasting model, its capacity feature set may include processed monthly chemical product production index changes, industry average capacity utilization rate changes, and other sequences. The data in the capacity feature set is strictly aligned with the aforementioned electricity load sequence at specific time points.

[0038] Step S3: Based on the power load sequence and the capacity feature set, identify the industry status type and calculate the status probability to generate a status probability sequence.

[0039] Step S3 of this application automatically identifies the potential economic fluctuation state of the industrial production system using statistical methods and quantifies the probability of belonging to different states at each point in time. Specifically, the data obtained in step S2 includes a production electricity load sequence that, after temperature effect correction, more purely reflects the intensity of production activities; and a capacity feature set containing stationarization indicators that directly measure production outcomes, such as changes in the industrial output index. These two sequences are precisely aligned in time, forming the basis for observing the dynamics of the economic system. The goal of step S3 is to discover industrial states within this system that are not directly observable but possess different statistical characteristics, such as a "low-volatility normal expansion period" and a "high-volatility crisis recession period," and to generate a state probability sequence containing continuous probability and discrete type identifiers, providing crucial input for subsequent state-dependent predictions.

[0040] First, the electricity load sequence and capacity feature set are grouped. The preset time point set used in this application refers to the entire time index of the data, such as each month from January 2006 to December 2023. The grouping operation of this application binds the electricity load data point corresponding to each time point (e.g., March 2020) with the capacity feature data point (e.g., the month-on-month change in industrial output for that month), forming an independent state identification sample group. Therefore, the obtained state identification sample group is a series of samples arranged in chronological order, where each sample contains a set of "feature-target" pairs: electricity load as the feature variable and capacity feature as the target variable. The above grouping operation prepares structured data for subsequent analysis, ensuring that the economic situation at each time point can be described by an independent data unit.

[0041] Since the true state of industry in an economic context is implicit and initially unknown, this application employs a pre-defined random assignment method to assign an initial industry state type label to each time point's state identification sample group. The industry state type is a predefined classification, such as simplified as "State 1" and "State 2," representing two hypothetical economic fluctuation patterns (e.g., high-volatility industry state and low-volatility industry state). The random assignment method used in this application's embodiments could be a simple coin toss, randomly assigning a label of "State 1" or "State 2" to each month, thereby generating an initial sequence of industry state types. This sequence is merely an initial, random guess used to start the algorithm, its significance being to provide a rough starting point for the system, allowing for subsequent iterative optimization and correction through data-driven methods. For example, when analyzing data during the 2008 financial crisis, the initial random label might incorrectly label September 2008 (when Lehman Brothers collapsed) as "State 1," and the subsequent task of the algorithm is to correct such errors.

[0042] Based on this initial randomized label sequence, further statistical analysis is needed to initialize key parameters. First, a time-series statistical analysis of the relationship between production capacity and electricity load is performed on all state-identified sample groups under the same industry state type. Specifically, sample data belonging to the same state label (e.g., all months labeled "State 1") are extracted, and linear regression analysis is performed on the production capacity characteristics (dependent variable) and electricity load (independent variable) of these data. The slope, intercept, and residual variance obtained from the regression analysis constitute the production capacity and load correlation index corresponding to that industry state type. The production capacity and load correlation index quantifies the average dependence of industrial production changes on electricity consumption changes under a specific assumed state. For example, for the initial sample group of "State 1," the regression may yield a smaller slope, indicating that the marginal demand for electricity from output is relatively stable under this state; while for the initial sample group of "State 2," the regression may yield a larger residual variance, suggesting greater volatility in economic relations under this state. Second, the transition frequency of the initial industry state type sequence is statistically analyzed according to time sequence. This operation examines the transitions between state labels at adjacent time points in the industry state type sequence. For example, it counts the number of events such as transitioning from "State 1" to "State 2" and vice versa. Dividing these transition counts by the total number of times the state occurs yields the initial state transition probabilities. For instance, a higher probability of transitioning from "State 1" to "State 2" indicates that the state is persistent. The capacity-load correlation indicators and state transition probabilities obtained from the above operation together constitute an initial, yet-to-be-optimized, hidden Markov model framework.

[0043] Subsequently, using a pre-defined likelihood function algorithm, the expected probability of each time point belonging to each industry state type is calculated based on the latest production capacity and load correlation indicators and state transition probabilities. The likelihood function is built upon a Hidden Markov Model (HMM), which calculates the probability of observing the entire data sequence given current parameters and decomposes it using a forward-backward algorithm. In specific calculations, the algorithm uses current correlation indicators to assess which state's statistical relationship the data at each time point better conforms to, while also considering the continuity of the state sequence in conjunction with transition probabilities. The calculated expected probability is a value between 0 and 1, representing the probability that the time point belongs to a certain state, and for each time point, the sum of the probabilities of belonging to all states is 1. For example, after the first likelihood function algorithm calculation, the expected probability of the time point in September 2008 belonging to "State 2" (high volatility) might increase from the initial 0.5 (random guess) to 0.7, because its drastic changes in output and load better match the high variance characteristic currently defined in the "State 2" model.

[0044] Next, the capacity-load correlation indicators and state transition probabilities of the corresponding time point state identification sample groups are updated using a preset maximization algorithm based on the expected probabilities obtained from the above operations. Specifically, when updating the correlation indicators, a weighted least squares regression analysis is performed again, where the weight of each time point sample is its expected probability of belonging to that state. This makes the parameter estimation of the most likely state to which the data point belongs more significant. Simultaneously, when updating the state transition probabilities, instead of statistically converting hard labels, the "expected" number of transitions is calculated based on the expected probabilities. For example, when calculating the probability of transitioning from "state 1" to "state 2", the numerator is the sum of the products of the expected probabilities of being in "state 1" and "state 2" in the two adjacent months, and the denominator is the sum of the expected probabilities of being in "state 1" in the first month. By providing the above maximization operation, the key parameters (i.e., capacity-load correlation indicators and state transition probabilities) are optimized based on the current probability perception.

[0045] To evaluate the convergence of this iterative optimization, this application needs to calculate the first change in the capacity and load correlation indicators before and after the update, and the second change in the state transition probability. The changes in this application are calculated as the Euclidean distance or relative rate of change between parameter vectors. Furthermore, the calculated first and second changes are compared with preset change thresholds. If either change is not less than the threshold, it means the model parameters are not yet stable, and the process of "likelihood function algorithm calculation - maximization algorithm calculation - change calculation before update" needs to be repeated until both the first and second changes are less than the preset change thresholds, indicating that the model parameters are sufficiently stable and the iterative optimization is complete. At this point, the expected probability calculated by the last likelihood function algorithm is considered the final reliable inference and is used as the state probability at the corresponding time point.

[0046] After iteration, this application determines the industry state type for the corresponding time point based on a preset probability threshold (usually set to 0.5) and the final determined state probability. For a given time point, if the probability of belonging to "State 2" (high volatility) is greater than 0.5, then that time point is classified as a "high volatility state type"; otherwise, it is classified as a "low volatility state type". Threshold comparison transforms continuous probability variables into clear classification labels, facilitating understanding and application. For example, in the analysis results, October 2008 might be assigned a state probability of 0.95 (belonging to a high volatility state). Since 0.95 > 0.5, the industry state type for that month is clearly determined to be "high volatility".

[0047] Finally, the state probabilities at all time points and the resulting industry state types are arranged and combined according to chronological order to generate the final required state probability sequence. This sequence is a data structure arranged in chronological order, where each row contains a time point, the probability value of each state, and the final determined state type. For example, the record for September 2008 in the sequence might be: State Probability (State 2) = 0.78, Industry State Type = "High Volatility". This state probability sequence fully characterizes the volatile state and uncertainty of the economic system over time. As a key feature input, it enables subsequent prediction models to dynamically adjust their prediction logic based on the identified economic environment (such as normal periods or crisis periods), thereby significantly improving the predictive model's adaptability to different economic scenarios and overall prediction accuracy.

[0048] The operation of step S3 above achieves the core objective of automatically learning and quantifying unobservable economic states from observable economic and power data through a complete process from random initialization to iterative convergence.

[0049] Step S4: Integrate the production power load sequence and the capacity feature set according to the state probability sequence to obtain the prediction variable set for each industry state type.

[0050] To address the challenge of effectively integrating frequently updated electricity data with low-frequency output indicators, and ensuring that this integration dynamically reflects different fluctuations in the economic system, step S4 of this application integrates the data generated in steps S2 and S3 to obtain a multi-frequency data set (i.e., a set of predictive variables). The predictive variable set's internal structure clearly distinguishes different industry state types, thus providing directly usable, information-rich, and frequency-consistent input features for subsequently constructing state-dependent predictive equations.

[0051] It should be understood that the state probability obtained from the above operations specifically refers to the probability value of each point in time belonging to a certain dominant industry state type (e.g., "high volatility state"). It is a continuous variable between 0 and 1, quantifying the degree of certainty about the current economic environment. The higher the probability that a point in time is judged as "high volatility state," the more significantly the characteristic pattern contained in the data at that moment deviates from the normal. When integrating data, assigning weights to such data points that match their state probabilities enables the integrated features to more sensitively capture and amplify the signal differences under different states. For example, during the deepening of the 2008 financial crisis, if the probability of a certain week's data being assigned to a "high volatility state" was 0.95, then the change in electricity load for that week would receive a very high weight of 0.95 in subsequent integration; conversely, during a period of economic stability, a week's data might only receive a "high volatility" weight of 0.1. In this way, the state probability is transformed from a descriptive indicator into a regulatory parameter guiding data synthesis (i.e., the integration weight coefficient).

[0052] Next, based on the preset frequency period and the aforementioned fusion weighting coefficients, the load values ​​in the production power load sequence are weighted and fused to obtain a composite power load sequence. It should be understood that the frequency period defines the target frequency for data aggregation, typically a rule for aggregating high-frequency data (e.g., weekly) to a low-frequency target (e.g., monthly). Production power load sequences usually contain observations at higher frequencies than the capacity feature set (e.g., 4 to 5 weeks of weekly data within a month). Simple monthly averaging loses information about inter-weekly variations and ignores the actual differences in the contribution of different weeks to the total monthly output. Therefore, the weighted fusion operation used in this application is not a simple summation or averaging, but a refined aggregation that considers time and state weights. Specifically, for a target month, all weekly data points belonging to that month need to be identified. The original load change value of each weekly data point is first multiplied by its corresponding state probability weighting coefficient for the first weighting, to amplify or suppress the contribution of that week's data in the monthly integration. Secondly, the time weight of the weekly data within the month also needs to be considered. This time weight can be determined based on the number of calendar days in the week or a preset working day equivalent to ensure that the contribution of weeks spanning the month is reasonably allocated. Ultimately, after these two weight adjustments, all weekly load change values ​​for the month are aggregated into a single monthly load change index, i.e., a composite power load sequence, through weighted averaging or weighted summation. For example, for October 2008, the four weeks of data included might be assigned high state probability weights of 0.92, 0.95, 0.97, and 0.96 respectively. Combined with the actual number of days in each week, a weighted fusion yields a monthly composite load value that strongly reflects the drastic fluctuations during the crisis. The frequency of the obtained composite power load sequence is now completely consistent with the monthly capacity feature set, but its connotation has surpassed the original monthly load; it condenses the time-series patterns and state information in high-frequency power data.

[0053] Finally, based on the same frequency cycle and the industry state type directly determined or derived from the state probability sequence, the composite power load sequence, the state probability sequence itself, and the capacity feature set are grouped and integrated to obtain the predictive variable set for each industry state type. First, all sequences (composite load sequence, state probability, and capacity features) are time-aligned according to monthly frequency. Next, data is integrated according to the industry state type. Specifically, based on the final determined industry state type for each month (e.g., "low volatility" or "high volatility"), the data for the entire time span is divided into two subsets. For each state type subset, such as "high volatility state type," its corresponding predictive variable set will contain three key pieces of information: first, the "composite power load" for all months under this state, serving as the core predictor; second, the "state probability values" corresponding to these months, which can serve as continuous control variables or the basis for interaction terms, used to characterize the confidence gradient within the state; and third, other relevant variables selected from the "capacity feature set" for the same month, such as quarterly dummy variables and previous output changes, serving as auxiliary control variables. For example, the predictor variable set constructed for the "high volatility state type" might contain data rows for September, October, and November of 2008. Each row contains the high-weighted composite load value, a state probability value close to 1.0, and other economic control variables for that month. Through the above grouping and integration operations, the final output is two or more independent sets of predictor variables bound to the state.

[0054] The operation of step S4 above creates the conditions for constructing independent prediction equations for "low volatility" and "high volatility" states, enabling the prediction model to use differentiated parameters to characterize the relationship between electricity consumption and industrial output under different economic environments. For example, in a crisis state, the model can learn a larger load factor to reflect the characteristic that output is more sensitive to changes in electricity during that period.

[0055] Step S5: Construct an industrial production prediction model based on the state probability sequence and the set of prediction variables.

[0056] First, the set of predictive variables generated in step S4 above is directly used as the prediction sample dataset for the corresponding industry state type. This prediction sample dataset contains several monthly samples, with each sample row including the core predictor for that month (e.g., composite electricity load), state identification information (e.g., state probability value), and other control variables (e.g., quarterly dummy variables, historical output inertia terms, etc.). Using this directly as the sample dataset defines a clear and homogeneous data range for subsequent statistical modeling. For example, when constructing a prediction equation for a "high volatility state," the training data used will be strictly limited to those historical months identified as high-volatility, ensuring that the patterns learned from this data specifically characterize the behavior of the economic system under stress and uncertainty. This approach of isolating samples by state is based on a key understanding: the production function and factor response elasticity of the economy may exhibit structural differences between normal and crisis periods. Fitting a single equation with all data would obscure these differences, causing the model to fail at inflection points.

[0057] Subsequently, a pre-defined steady-state estimation regression algorithm is used to perform independent regression analysis on each divided prediction sample dataset. The steady-state estimation regression algorithm used in this application specifically refers to a class of parameter estimation methods with strong robustness, such as M-estimation based on Huber loss or quantile regression. This type of algorithm is chosen instead of ordinary least squares because, although the data has undergone the aforementioned complex processing to eliminate major disturbances, unpredictable sporadic events may still occur in industrial production activities, manifesting as statistical outliers in the data. If these outliers are not controlled, they will excessively influence parameter estimation, impairing the model's general representativeness. Robust regression algorithms, by reducing the weight of outlier residuals, ensure that the obtained parameter estimates more stably reflect the dominant and persistent data patterns under that state. The specific form of the regression analysis is to construct a linear regression model for a state sample dataset, using industrial output change as the explained variable, composite electricity load as the core explanatory variable, and incorporating other control variables. Solving this model using the steady-state estimation algorithm yields a set of coefficient estimates, intercept terms, and error variance estimates, i.e., variable prediction correlation data. For example, regression analysis on a sample set of "high-fluctuation conditions" may yield significantly larger coefficients for the core load variable than those obtained from a sample set of "low-fluctuation conditions." Variable predictive correlation data quantitatively reveal, numerically, the sensitivity of industrial production changes to changes in electricity consumption (i.e., load factor), inherent growth or contraction trends (intercept term), and the level of predictive uncertainty (error variance) under different economic conditions. It paints a specific numerical portrait of each implied economic condition.

[0058] Finally, based on the obtained variable prediction correlation data and the corresponding prediction sample dataset, regression equation models for each industry state type are constructed and integrated to obtain a complete industrial production prediction model. The numerical parameters obtained in the previous step (i.e., variable prediction correlation data) are transformed into mathematical expressions that can be directly used for prediction. For each industry state type, the standard form of its regression equation model can be expressed as: Monthly industrial output change forecast = (intercept term in this state) + (load coefficient in this state) × (composite electricity load in the current month) + (sum of the products of the coefficients of other control variables and their values) + random error term. For example, the equation constructed for "low volatility state" may be in the form of: Output change = 0.2 + 0.65 × composite load + 0.1 × quarterly effect adjustment; while the equation constructed for "high volatility state" may be in the form of: Output change = -0.5 + 0.95 × composite load + 0.15 × quarterly effect adjustment. These two equations intuitively demonstrate state dependence: in a crisis state, the same increase in electricity load may predict a larger rebound in output (load coefficient 0.95 > 0.65), but it is also accompanied by a greater downside risk (negative intercept term). Combining multiple regression equations constructed separately for different states forms an industrial production forecasting model. The operating logic of this forecasting model is dynamic: when a forecast for a new month is needed, the system first calculates the real-time probability of the month belonging to each state based on the latest high-frequency electricity data, using the logic of step S3 above (or using the estimated state transition probabilities). Subsequently, the system does not rigidly select one equation, but rather weights and combines the forecast results of multiple equations based on these real-time probabilities. For example, if the probability of a month being classified as a high-fluctuation state is 0.8 and the probability of being classified as a low-fluctuation state is 0.2, then the final forecast value = 0.8 × (high-fluctuation equation forecast value) + 0.2 × (low-fluctuation equation forecast value). This probability-based soft-switching mechanism enables the entire forecasting system to smoothly adapt to gradual changes and transitions in economic conditions, avoiding jumps in predicted values ​​at state boundaries while preserving the model's ability to cope with uncertainty. Therefore, step S5 does not produce a static formula, but rather an intelligent forecasting system that integrates state identification results, possesses parameter heterogeneity, and features dynamic weighting logic. It encapsulates the patterns revealed in the data under different economic systems into computable modules, providing a complete mathematical model foundation for ultimately conducting accurate and environmentally adaptive industrial production forecasts.

[0059] This application applies to the field of Internet of Things (IoT) technology. It generates a time-series variation sequence by periodically decomposing and smoothing a standard industrial dataset according to a time period. The time-series variation sequence is then segmented and its power load is adjusted and features extracted to obtain a production power load sequence and a capacity feature set. Based on the production power load sequence and capacity feature set, industry state types are identified and state probabilities are calculated to generate a state probability sequence. The production power load sequence and capacity feature set are then integrated based on the state probability sequence to obtain a set of predictive variables for each industry state type. An industrial production prediction model is then constructed based on the state probability sequence and the predictive variable set. This application improves the accuracy, robustness, and interpretability of near-term industrial output prediction by constructing a state conditional regression model through temperature and calendar separation, frequency mixing fusion, and state identification.

[0060] like Figure 2 The diagram shown is a functional block diagram of an industrial production prediction model construction device provided in an embodiment of this application.

[0061] In some embodiments, the industrial production forecasting model building apparatus 2 may include multiple functional modules composed of computer program segments. The computer programs for each program segment in the industrial production forecasting model building apparatus 2 may be stored in the server's memory and executed by at least one processor to perform (see details). Figure 1 (Description) Functionality of the method for constructing industrial production forecasting models.

[0062] In this embodiment, the industrial production prediction model construction device 2 can be divided into multiple functional modules according to its functions. These functional modules may include: a data optimization module 21, a decomposition and smoothing module 22, a feature extraction module 23, a state analysis module 24, a variable integration module 25, and a model construction module 26. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module will be detailed in subsequent embodiments.

[0063] Data optimization module 21 is used to preprocess the received industrial dataset to obtain a standard industrial dataset; The decomposition and smoothing module 22 is used to perform periodic decomposition and trend smoothing on the standard industrial dataset according to a preset time period to generate a time-series change sequence. The feature extraction module 23 is used to perform segmented adjustment and feature extraction on the time-series change sequence of power load to obtain the production power load sequence and the capacity feature set; The state analysis module 24 is used to identify the industry state type and calculate the state probability based on the production power load sequence and the capacity feature set, and generate a state probability sequence. The variable integration module 25 is used to integrate the production power load sequence and the capacity feature set according to the state probability sequence to obtain a set of predictive variables for each industry state type. The model building module 26 is used to build an industrial production prediction model based on the state probability sequence and the set of prediction variables.

[0064] In an optional implementation, the data optimization module 21 is specifically used for: The received industrial dataset is time-domain aligned to obtain a time-synchronized dataset; Based on the time sequence, the time-series synchronized dataset is subjected to abnormal data identification and cleaning, as well as interpolation compensation for missing data, to obtain a clean industrial dataset. The importance of the clean industrial dataset is assessed based on the preset production planning data to obtain the importance labels corresponding to each time point. The weight coefficients corresponding to each time point are obtained from the preset set of weight coefficients based on the importance markers; The data in the clean industrial dataset are weighted and corrected according to the weighting coefficients to obtain a standard industrial dataset.

[0065] In an optional implementation, the decomposition smoothing module 22 is specifically used for: The production capacity-related dataset is periodically decomposed according to a preset time period using a preset seasonal adjustment algorithm to obtain industrial production capacity data sequences for each time period. The industrial capacity data sequence is processed by a preset differential algorithm to generate production trend data sequences for each time period. Heating and cooling demands are calculated based on preset base temperature values ​​and the ambient temperature dataset to obtain temperature regulation indices for each time period. The temperature regulation index and the production trend data sequence are aligned and merged according to the time sequence to generate a time-series change sequence.

[0066] In an optional implementation, the feature extraction module 23 is specifically used for: The non-production power load adjustment amount is calculated in each time period based on the temperature regulation index using a preset temperature and load correlation algorithm. Based on the power load adjustment amount, the production trend data sequence is segmented and corrected for power load to obtain the production power load sequence. The time-series change sequence is processed to extract industrial output change features to obtain a capacity feature set.

[0067] In an optional implementation, the state analysis module 24 is specifically used for: Step S31: Based on the preset time point set, group the production power load sequence and the production capacity feature set to obtain the state identification sample group for each time point. Step S32: Randomly label the state identification sample group with preset industry state types using a preset random allocation method to generate an industry state type sequence corresponding to each time point; Step S33: Perform time-series change statistical analysis on the state identification sample group of the same industry state type to obtain the capacity and load correlation index corresponding to each industry state type, and perform industry state type transition frequency statistics on the industry state type sequence according to the time sequence to obtain the transition probability between states. Step S34: Using a preset likelihood function algorithm, calculate the expected probability of each time point belonging to each industry state type based on the capacity and load correlation index and the state transition probability. Step S35: Update the capacity and load correlation index and state transition probability of the state identification sample group at the corresponding time point according to the expected probability using a preset maximization algorithm. Step S36: Calculate the first change in the capacity and load correlation index before and after the update, and calculate the second change in the state transition probability before and after the update; Repeat steps S34 to S36 until both the first change and the second change are less than the preset change threshold, and use the expected probability as the state probability at the corresponding time point. Step S37: Determine the industry state type at the corresponding time point based on the preset probability threshold and the state probability; Step S38: Arrange and combine the state probabilities and the industry state types according to the time sequence to generate a state probability sequence.

[0068] In an optional implementation, the variable integration module 25 is specifically used for: The state probability is used as the fusion weighting coefficient of the corresponding load in the power production load sequence; Based on the preset frequency period and the fusion weighting coefficient, the load quantities in the power production load sequence are weighted and fused to obtain a composite power load sequence. Based on the frequency period and the industry status type, the composite power load sequence, the status probability sequence, and the capacity feature set are grouped and integrated to obtain a set of predictive variables for each industry status type.

[0069] In an optional implementation, the model building module 26 is specifically used for: The set of predictive variables corresponding to the industry status type is used as the prediction sample dataset for each industry status type. By using a preset steady-state estimation regression algorithm, regression analysis is performed on the predicted sample dataset to obtain variable prediction correlation data for each industry state type; Based on the predicted correlation data of the variables and the predicted sample dataset, regression equation models for each industry status type are constructed to obtain industrial production prediction models.

[0070] It should be understood that the various variations and specific embodiments of the methods provided in the above embodiments are also applicable to the industrial production prediction model construction apparatus of this embodiment. Through the foregoing detailed description of the industrial production prediction model construction method, those skilled in the art can clearly understand the implementation method of the industrial production prediction model construction apparatus of this embodiment. For the sake of brevity, it will not be described in detail here.

[0071] like Figure 3 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this application.

[0072] In a preferred embodiment of the present invention, the electronic device 3 may include, but is not limited to, a memory 31, at least one processor 32, and at least one communication bus 33.

[0073] Those skilled in the art should understand that Figure 3 The structure of the electronic device 3 shown does not constitute a limitation of the embodiments of the present invention. The electronic device 3 may also include more or fewer other hardware or software than shown, or different component arrangements.

[0074] In some embodiments, the electronic device 3 is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits, programmable gate arrays, digital processors, and embedded devices.

[0075] It should be noted that the electronic device 3 is merely an example. Other existing or future electronic products that are suitable for this application should also be included within the scope of protection of this application and are incorporated herein by reference.

[0076] In some embodiments, the memory 31 stores a computer program that, when executed by the at least one processor 32, implements all or part of the steps in the method for constructing the industrial production prediction model as described above. The memory 31 includes read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data. Further, the computer-readable storage medium may primarily include a program storage area and a data storage area, wherein the program storage area may store an operating system, at least one application program required for a function, etc.

[0077] In some embodiments, the at least one processor 32 is the control unit of the electronic device 3, connecting various components of the electronic device 3 via various interfaces and lines. It executes programs or modules stored in the memory 31 and calls data stored in the memory 31 to perform various functions and process data. For example, when the at least one processor 32 executes a computer program stored in the memory 31, it implements all or part of the steps of the industrial production prediction model construction method described in this application embodiment; or it implements all or part of the functions of the industrial production prediction model construction device. The at least one processor 32 may be composed of integrated circuits, such as a single-packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips.

[0078] In some embodiments, the at least one communication bus 33 is configured to enable communication between the memory 31 and the at least one processor 32, etc. Although not shown, the electronic device 3 may also include a power supply (e.g., a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 32 via a power management device, thereby enabling functions such as charging, discharging, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device 3 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0079] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause an electronic device (which may be a personal computer, electronic device, or network device, etc.) or processor to execute portions of the methods described in the various embodiments of this application.

[0080] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0081] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for constructing an industrial production prediction model, characterized by, The method includes: The received industrial dataset is preprocessed to obtain a standard industrial dataset, which is then periodically decomposed and trend smoothed according to a preset time period to generate a time-series change sequence. The power load is segmented and its features are extracted from the time-series variation sequence to obtain the power load sequence and capacity feature set; Based on the power load sequence and the capacity feature set, the industry status type is identified and the status probability is calculated to generate a status probability sequence. Based on the state probability sequence, the production power load sequence and the capacity feature set are integrated to obtain a set of predictive variables for each industry state type; An industrial production prediction model is constructed based on the state probability sequence and the set of prediction variables.

2. The method of claim 1, wherein, The process of preprocessing the received industrial dataset to obtain a standard industrial dataset includes: The received industrial dataset is time-domain aligned to obtain a time-synchronized dataset; Based on the time sequence, the time-series synchronized dataset is subjected to abnormal data identification and cleaning, as well as interpolation compensation for missing data, to obtain a clean industrial dataset. The importance of the clean industrial dataset is assessed based on the preset production planning data to obtain the importance labels corresponding to each time point. The weight coefficients corresponding to each time point are obtained from the preset set of weight coefficients based on the importance markers; The data in the clean industrial dataset are weighted and corrected according to the weighting coefficients to obtain a standard industrial dataset. 3.The method of claim 1, wherein the standard industrial dataset comprises a production capacity correlation dataset and an ambient temperature dataset. The step of performing periodic decomposition and trend smoothing on the standard industrial dataset according to a preset time period to generate a time-series change sequence includes: The production capacity-related dataset is periodically decomposed according to a preset time period using a preset seasonal adjustment algorithm to obtain industrial production capacity data sequences for each time period. The industrial capacity data sequence is processed by a preset differential algorithm to generate production trend data sequences for each time period. Heating and cooling demands are calculated based on preset base temperature values ​​and the ambient temperature dataset to obtain temperature regulation indices for each time period. The temperature regulation index and the production trend data sequence are aligned and merged according to the time sequence to generate a time-series change sequence.

4. The method of claim 3, wherein, The step of segmenting and extracting features from the time-series variation sequence of power load to obtain the production power load sequence and capacity feature set includes: The non-production power load adjustment amount is calculated in each time period based on the temperature regulation index using a preset temperature and load correlation algorithm. Based on the power load adjustment amount, the production trend data sequence is segmented and corrected for power load to obtain the production power load sequence. The time-series change sequence is processed to extract industrial output change features to obtain a capacity feature set.

5. The method of claim 1, wherein, The step of identifying industry state types and calculating state probabilities based on the power load sequence and the capacity feature set, and generating a state probability sequence, includes: Step S31, grouping the production power load sequence and the set of capacity characteristics according to a set of preset time points to obtain a state recognition sample group at each time point; Step S32, randomly labeling the state recognition sample group with a preset industrial state type through a preset random allocation method to generate an industrial state type sequence corresponding to each time point; Step S33, performing time sequence change statistical analysis between capacity and production power load on the state recognition sample group of the same industrial state type to obtain a capacity and load correlation index corresponding to each industrial state type, and performing industrial state type transition frequency statistics on the industrial state type sequence according to time sequence to obtain a state transition probability; Step S34, calculating an expected probability of each time point belonging to each industrial state type according to the capacity and load correlation index and the state transition probability through a preset likelihood function algorithm; Step S35, updating the capacity and load correlation index and the state transition probability of the state recognition sample group at the corresponding time point according to the expected probability through a preset maximization algorithm; Step S36, calculating a first change amount of the capacity and load correlation index before and after updating, and calculating a second change amount of the state transition probability before and after updating; Repeat the steps S34 to S36 until the first change amount and the second change amount are both less than a preset change amount threshold, and take the expected probability as a state probability of the corresponding time point; Step S37, determining an industrial state type of the corresponding time point according to a preset probability threshold and the state probability; Step S38, arranging and combining the state probability and the industrial state type according to time sequence to generate a state probability sequence.

6. The method of claim 5, wherein, The integrating the production power load sequence and the set of capacity characteristics according to the state probability sequence to obtain a set of predicted variables of each industrial state type includes: Taking the state probability as a fusion weight coefficient of the corresponding load in the production power load sequence; Weighting and fusing each load in the production power load sequence according to a preset frequency period and the fusion weight coefficient to obtain a composite power load sequence; Grouping and integrating the composite power load sequence, the state probability sequence and the set of capacity characteristics according to the frequency period and the industrial state type to obtain a set of predicted variables of each industrial state type.

7. The method of claim 5, wherein the step of constructing the industrial production prediction model is performed by using a machine learning algorithm. The constructing an industrial production prediction model according to the state probability sequence and the set of predicted variables includes: Taking the set of predicted variables corresponding to the industrial state type as a prediction sample data set of each industrial state type; Performing regression analysis on the prediction sample data set through a preset steady-state estimation regression algorithm to obtain variable prediction correlation data of each industrial state type; Performing regression equation model construction of each industrial state type according to the variable prediction correlation data and the prediction sample data set to obtain an industrial production prediction model.

8. An apparatus for constructing an industrial production prediction model, applied to the method for constructing an industrial production prediction model according to claim 1, characterized in that, The device includes: A data optimization module configured to preprocess a received industrial data set to obtain a standard industrial data set; The decomposition and smoothing module is configured to periodically decompose and trend smooth the standard industrial dataset according to a preset time period, to generate a time-series variation sequence; The feature extraction module is configured to segment and adjust the time-series variation sequence, and extract features, to obtain a production power load sequence and a production capacity feature set; The state analysis module is configured to identify an industrial state type and calculate a state probability according to the production power load sequence and the production capacity feature set, to generate a state probability sequence; The variable integration module is configured to integrate the production power load sequence and the production capacity feature set according to the state probability sequence, to obtain a prediction variable set of each industrial state type; The model construction module is configured to construct an industrial production prediction model according to the state probability sequence and the prediction variable set.

9. An electronic device, comprising: The electronic device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the construction method of the industrial production prediction model according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the construction method of the industrial production prediction model according to any one of claims 1 to 7.

Images