Periodic settlement data prediction method fusing nonlinear dynamic lag modeling

Abstract

The invention discloses a periodic settlement data prediction method fused with nonlinear dynamic lag modeling, which comprises the following steps: counting the number of manual operation times in unit time as an intervention frequency, comparing the intervention frequency with a preset threshold value, and judging whether a service scene is a periodic scene or a non-periodic scene; for the scene which is determined to be periodic, calibrating a monthly prediction result based on scale change of historical same-period data; for the scene judged to be aperiodic, parameters are updated in combination with a moving average method, and processing delay is dynamically adjusted by incorporating holidays and festivals and resting arrangement, so that sequence prediction is generated and completed; and dynamically updating model parameters and predicted values. According to the method, manual intervention intensity and sequence period feature analysis are fused, the universality and accuracy of service scene judgment are remarkably improved, the recognition limitation of a traditional method on mixed services is broken through, and lagging behavior prediction in the fields of energy, finance and the like can be adapted through the three stages of triggering, processing and completing.

Description

Technical Field

[0001] This invention relates to a method for predicting periodic settlement data that integrates nonlinear dynamic lag modeling, belonging to the field of time series modeling and complex behavior prediction technology, and is particularly suitable for multi-stage lag behavior modeling and numerical flow trend prediction in periodic resource settlement systems. Background Technology

[0002] Existing forecasting techniques mostly focus on single-stage or simple lag scenarios (such as sales forecasting and inventory turnover), while research on modeling multi-stage, strongly correlated lag behaviors (such as "trigger-response-completion") is relatively limited. Although such problems are widespread in fields such as energy, finance, and logistics (such as "issuance-payment-receipt" in expense cash flow and "order-shipment-receipt" in logistics), existing methods often suffer from insufficient forecasting accuracy because they ignore the dynamic correlation between stages or external interference factors (such as holidays and policy changes).

[0003] ① Traditional methods (such as ARIMA and regression analysis) struggle to capture the nonlinear relationships and dynamic lag patterns of multi-stage behaviors. In cost cash flow forecasting, a threshold effect may exist between user payment behavior and issuance volume, but the ARIMA model, limited to linear relationships, cannot capture such nonlinear responses. Furthermore, traditional methods assume fixed lag patterns (such as the fixed order in ARIMA), but in real-world scenarios, the "issuance-payment" delay may dynamically change with user type or seasonal fluctuations, leading to significant prediction biases.

[0004] ② While machine learning models (such as neural networks) can handle complex relationships, they suffer from poor interpretability and rely heavily on labeled data. As "black box" models, neural networks struggle to provide business-understandable decision-making support. For example, when key driving factors need clear explanation, the feature importance analysis of neural networks is often ambiguous. Furthermore, supervised learning requires a large amount of high-quality labeled data, but in emerging fields or long-tail scenarios, data sparsity can lead to model overfitting or failure. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] In view of the problems existing in the above and / or existing periodic settlement data prediction methods that integrate nonlinear dynamic hysteresis modeling, this invention is proposed.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: The first invention proposes a method for predicting periodic settlement data that integrates nonlinear dynamic hysteresis modeling, characterized by comprising the following steps: The number of manual operations on business data per unit time in the target business scenario is used as the intervention frequency. The intervention frequency is compared with a preset threshold. When the intervention frequency is lower than the preset threshold, a monthly periodic judgment is made to determine whether the business scenario is a periodic scenario or a non-periodic scenario. The settlement behavior of the target business scenario is broken down into the triggering stage, the processing stage, and the completion stage, and correspondingly associated with the historical triggering time series data, historical processing time series data, and historical completion time series data in the business data. For periodic scenarios, a time series prediction model is used to predict the completed time series data, and the prediction results of the completed time series data are calibrated by scaling transformation based on historical data from the same period. For non-periodic scenarios, the weights from the trigger stage to the processing stage and from the processing stage to the completion stage are dynamically allocated based on historical business data. The parameters are updated using the moving average method, and the processing delay is dynamically adjusted by incorporating holidays and work schedule adjustments, in order to generate prediction results for the completion time series data. Output the prediction results of the completed time series data, and dynamically update the model parameters and prediction values ​​as new data arrives.

[0008] In a preferred embodiment, the step of performing monthly periodic determination includes: Let the completion time series be (x1, x2, ..., x...). n ), k=[n / 30], then let X be... i = (x i+1 x i+2， ...,x i+30 (), where k is the monthly coefficient, and n represents the nth day of the month. The monthly cycle correlation coefficient is defined as: ; Calculate the monthly subsequence correlation of historical completed sequence data for the target business scenario. When the correlation coefficient exceeds a preset threshold, monthly periodicity is confirmed and the scenario is determined to be periodic; otherwise, it is a non-periodic scenario.

[0009] In a preferred embodiment, the step of dividing the target settlement behavior into a triggering phase, a processing phase, and a completion phase includes: The triggering phase corresponds to a business initiation event, i.e., a bill generation operation; the processing phase corresponds to a user response event, i.e., a fund transfer operation; and the completion phase corresponds to a fund settlement event, i.e., a settlement completion operation.

[0010] In a preferred embodiment, the time series forecasting model makes predictions by decomposing trend terms, seasonal terms, and holiday terms; The time series forecasting model is expressed by the following formula:

[0011] Where g(t) is the trend component of the data, h(t) is the holiday component model, s(t) is the seasonal component model, and εt is the exogenous variable; Exogenous variables are introduced as inputs to the time series forecasting model, including policy adjustment points and system upgrade time points; The seasonal model is constructed using Fourier series and dynamically adjusts the period length to adapt to the monthly fluctuation patterns of different business scenarios.

[0012] In a preferred embodiment, the step of completing the prediction result of time series data by scaling calibration based on historical contemporaneous data specifically includes: Scale the time series data predicted by the trained time series model for the next month: ; Among them, W′ K ={W′ 1，K , W′ 2，K , ..., W′ tk，K} and W′ K+1 ={W′ 1，K+1 , W′ 2，K+1 , ..., W′ tk+1 , K+1} represent the completed time series data for month K and month K+1 of last year, respectively. K ={W 1，K W 2，K , ..., W tk，K} represents the completed time series data for month K of this year, Y K ={Y 1，K Y 2，K , ..., Y tk，K} represents the predicted sequence obtained in the previous step, Y i,K The data is based on historical data and scaled, where j is the j-th day of the K-th month.

[0013] As a preferred embodiment, the modeling of non-periodic behavior includes: quantifying the weight allocation rules of the triggering stage, processing stage, and completion stage based on historical data, and updating parameters in combination with moving averages; The triggering phase prediction includes: Calculate the moving average based on the sum of the monthly trigger sequences for the previous three months; Based on the daily distribution pattern of the trigger sequence in the same period last year, generate the daily trigger volume allocation weight for the next month; The estimated value of the trigger time series data is obtained by multiplying the moving average value by the trigger amount weight; The processing stage prediction includes: Set the processing weight for the current day, the processing weight for delayed working days, and the weight allocated over a period; The triggered time series data forecast values ​​are allocated to the corresponding processing days according to three weight categories to obtain the processed time series forecast values, where: The weighting of processing on the current day will be allocated to the next day. The weighting of delayed working days is allocated to the 5th working day; The remaining value is evenly distributed among the working days during the period; The completion stage prediction includes: Three types of delay weights are calculated based on historical data from the previous three months: weight for completion on the same day, weight for completion on the next day, and weight for delay due to holidays. The processing time series estimates are allocated to the corresponding completion dates according to lag weights to obtain the completion time series estimates, where: Weight allocation is completed and carried over to the same day; The weight allocation will be completed the following day and carried over to the next day. The weighting of holidays will be postponed to the first non-holiday period.

[0014] In a preferred embodiment, the dynamic updating of model parameters and predicted values ​​when new data arrives includes: When actual business data arrives, the dynamic weight allocation parameters are recalculated; Update the moving average using a rolling time window; Perform business rule verification on the prediction results, and trigger manual review when the deviation exceeds the tolerance threshold.

[0015] Secondly, the present invention provides an electricity cost prediction method, comprising the following steps: Collect historical electricity bill datasets; Based on the periodic settlement data prediction method that integrates nonlinear dynamic lag modeling as described in the first aspect above, multiple business scenarios in the historical electricity bill dataset are divided into periodic scenarios and non-periodic scenarios, and the daily electricity bill prediction data for the following month is obtained.

[0016] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a storage device having stored one or more programs thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any of the implementations of the first aspect above.

[0017] Fourthly, the present invention provides a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0018] This invention significantly improves the universality and accuracy of business scenario identification by integrating the intensity of human intervention with the analysis of sequence periodic characteristics, breaking through the limitations of traditional methods in identifying hybrid businesses. At the modeling level, it innovatively integrates external event response mechanisms and a multi-stage delay rule base, enhancing the robustness of periodic scenarios to policy adjustments while accurately capturing the complex patterns of cross-departmental collaboration in non-periodic processes, systematically solving prediction distortions caused by holiday backlogs and work schedule fluctuations. Based on a rolling update framework, it achieves closed-loop optimization of the prediction system, significantly reducing operational costs while maintaining real-time performance. This solution coordinates time series analysis and business process topology modeling with a unified architecture, setting a new industry benchmark in prediction accuracy, scenario coverage, and global support. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of a periodic settlement data prediction method that integrates nonlinear dynamic hysteresis modeling, as shown in Example 1.

[0021] Figure 2 This is a schematic diagram of the actual structure of a periodic settlement data prediction method that integrates nonlinear dynamic hysteresis modeling, as shown in Example 2.

[0022] Figure 3 This is a visualization of the overall trend of the simulated real expense receipt curve in Example 2, which is a periodic settlement data prediction method that integrates nonlinear dynamic hysteresis modeling. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] In the following description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways than those specifically described herein. Those skilled in the art may make similar extensions without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0025] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation manner of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or selectively exclusive embodiments from other embodiments.

[0026] Embodiment 1 Refer to Figure 1 This is the first embodiment of the present invention, which provides a method for predicting periodic settlement data by integrating non-linear dynamic lag modeling, including: The core idea of the general prediction framework is to abstract the multi-stage lag behavior into a three-stage process of "trigger → process → complete", and combine time series analysis with business logic modeling to achieve dynamic prediction and rolling update of the behavior of each stage. The time series prediction model also includes: Introduce exogenous variables as the input of the prediction model, and the external event variables include policy adjustment nodes and system upgrade time points; Construct seasonal components through Fourier series, and dynamically adjust the cycle length to adapt to the monthly fluctuation rules of different business scenarios.

[0027] ① Business scenario periodic determination framework based on intervention frequency The present invention proposes a simplified method for quantifying the degree of manual intervention. Only through this core index of intervention frequency, combined with the periodic characteristics of the historical data of the completion sequence, the dynamic determination of the periodicity / aperiodicity of the business scenario is realized. Define the intervention frequency as the number of manual operations per unit time, and the calculation formula is: (1) Through domain expert annotation, set the intervention frequency threshold f_threshold for the periodic scenario. When f_intervention < f_threshold, it is determined that the scenario is weakly affected by manual intervention and can enter the periodic detection process.

[0028] Record the completion time series as (x1, x2..., x n ), k = [n / 30], then record X i = (x i + 1, x i+2 ,... x i+30(), where k is the monthly coefficient and n is the n-day representation of the month, and the monthly cycle correlation coefficient is defined as: (2) ② Stage splitting and data mapping The target settlement behavior is broken down into a trigger phase, a processing phase, and a completion phase, including: The triggering phase corresponds to a business initiation event, i.e., a bill generation operation; the processing phase corresponds to a user response event, i.e., a fund transfer operation; and the completion phase corresponds to a fund settlement event, i.e., a settlement completion operation.

[0029] The target behavior of non-periodic business scenarios is broken down into a triggering phase (e.g., fee issuance), a processing phase (e.g., user payment), and a completion phase (e.g., fund arrival), with key parameters defined for each phase. Assuming the prediction time point is the end of month K, the prediction target is the completion time series for month K+1. The data used includes historical triggering time series, historical processing time series, and historical completion time series for each business scenario.

[0030] ③ Modeling of periodic behaviors (3.1) Prophet model For completed time series with significant periodicity, the Prophet model is used to decompose trend, seasonality, and holiday effects, and exogenous variables (such as policy nodes) are introduced to enhance predictive robustness. The model is expressed as follows: y(t)=g(t)+s(t)+h(t)+εt(3) g(t) represents the trend component of the data, used to fit the non-periodic changes in the data, describing how the sequence changes and how it is expected to continue changing. We typically use logistic growth to model this trend; its basic form is as follows: (4) s(t) represents the seasonality model. Human behavior represented by time series data in business analysis often exhibits seasonality over multiple periods. For example, weekdays and weekends may affect weekly recurring series, while seasons and holidays may affect yearly recurring series. Here, a flexible periodic effect model is simulated using Fourier series: (5) P is the period, a n The coefficient corresponds to the amplitude of the nth harmonic cosine wave, while b n The coefficient corresponds to the amplitude of the nth harmonic sine wave.

[0031] In the formula, h(t) represents the holiday term model. The Prophet model assumes that holidays have independent effects on the model; for the i-th holiday, let D... iThe function represents the date of the holiday, along with an index function indicating whether time t falls within holiday i and a parameter κ. i This indicates the extent of the holiday's impact. A regression matrix is ​​constructed in a manner similar to seasonality. Z(t)=[I(t∈D1),...,I(t∈DL)](6) Then we have h(t) = Z(t)κ, where the prior κ ~ Normal(0, ν2).

[0032] (3.2) Scale variation To ensure accuracy at the monthly level, the predicted sequence for the next month by the trained model is scaled, W′ K ={W′ 1，K , W′ 2，K , ..., W′ tk，K} and W′ K+1 ={W′ 1，K+1 , W′ 2，K+1 , ..., W′ tk+1 , K+1} represent the completion time series of month K and month K+1 of last year, respectively. K ={W 1，K W 2，K , ..., W tk，K} represents the completion time series for month K of this year, Y K ={Y 1，K Y 2，K , ..., Y tk，K} represents the predicted sequence obtained in the previous step, which is then scaled based on historical data. Let j be the j-th day of the k-th month. (7) ④ Modeling of non-periodic behavior (dynamic weight allocation) For non-periodic phases, the weighting pattern of triggering → processing → completion is based on historical data, combined with moving average parameter updates. Simultaneously, interference factors such as holidays and work schedule adjustments are incorporated to dynamically adjust processing delays.

[0033] (4.1) Predicting the trigger sequence for next month C K ={C 1，K C 2，K C tk，K}、C K−1 ={C 1，K−1 C 2，K−1 C tk−1，K−1} and C K−2 ={C 1，K−2 C 2，K−2 Ctk−2，K−2} represent the trigger time series for month K, month K+1, and month K+2 of this year, respectively. The moving average of the previous three months is used as the monthly sum of the trigger time series for the next month, i.e.: (8) When allocating the monthly sum of the trigger time series for the next month, the trigger time series from the same period last year should also be taken into consideration. K ′={C′ 1，K C′ 2，K ,...,C′ tk，K} represents the trigger time series for month K of last year. The trigger time series for the previous three months are also processed; taking last month's trigger sequence as an example.

[0034] (9) The allocation basis for calculating the monthly sum of the trigger sequence next month (w) 1，K+1 w 2，K+1 , ..., w tk+1，K+1 This allows the estimated value of the time series to be triggered next month to be obtained through C. i，K+1 =Cw i，K+1 The allocation is calculated as follows: (10) (4.2) Predict the processing sequence for next month Taking into account the impact of weekdays and weekends, the estimated values ​​of the daily trigger sequences are allocated using weights w1 and w2. w1C i，K+1 The processing sequence assigned to day i+1 will be w2C i，K+ 1. The processing sequence assigned to the 5th working day after day i, with (1−w1−w2)C i，K+1 The processing sequence is evenly distributed throughout the period to obtain the estimated value L of the processing sequence. K+1 ={L 1，K+1 L 2，K+1 , ..., L tk+1，K+1}

[0035] (4.3) Predict the completion sequence next month The i-th entry in the dataset is (y i u i v i ), where u i and v i Let a1 and a2 represent the processing and completion times corresponding to the value yi, respectively. Based on the data from the previous three months, calculate the "processing-completion" delay weights m1, m2, and m3. Let a1 = a2 = a3 = 0, and define a function f, if u i and v i If it is a holiday, then f(u)i ) = f(v i ) = 0, otherwise f(u) i ) = f(v i =1. The iterative processing for each data point is as follows: (11) u i to v i The date input during the period is 0. (12) Obtain the "processing-completion" delay weights m1, m2, and m3. Based on these, allocate the estimated processing time series values ​​for the following month, assigning m1L... i，K+1 The completed sequence assigned to day i will be m2L i，K+1 The completed sequence assigned to day i+1 will be m3L i，K+1 The completion sequence allocated to the first working day after day i is used to obtain the estimated value W of the completion sequence for the following month. K+1 ={W 1，K+1 W 2，K+1 , ..., W tk+1，K+1}

[0036] Modeling non-periodic behavior includes: quantifying the weight allocation patterns of the triggering, processing, and completion stages based on historical data, and updating parameters using moving averages; The triggering phase prediction includes: Calculate the moving average based on the monthly trigger sequence of the previous three months; Based on the daily distribution pattern of the trigger sequence in the same period last year, generate the daily trigger volume allocation weight for the next month; The estimated value of the trigger sequence is obtained by multiplying the moving average by the assigned weight; The processing stage prediction includes: Set the processing weight for the current day, the processing weight for delayed working days, and the weight allocated over a period; The trigger estimate will be allocated to the processing sequence according to the three weight categories mentioned above: The weighting of processing on the current day will be allocated to the next day. The weighting of delayed working days is allocated to the 5th working day; The remaining value is evenly distributed among the working days during the period; The completion stage prediction includes: Three types of delay weights are calculated based on historical data from the previous three months: weight for completion on the same day, weight for completion on the next day, and weight for delay due to holidays. The estimated values ​​will be allocated according to delay weights: Weight allocation is completed and carried over to the same day; The weight allocation will be completed the following day and carried over to the next day. The weighting of holidays will be postponed to the first non-holiday period.

[0037] Dynamically updating model parameters and predictions when new data arrives includes: When actual business data arrives, the dynamic weight allocation parameters are recalculated; Update the moving average using a rolling time window; Perform business rule verification on the prediction results, and trigger manual review when the deviation exceeds the tolerance threshold.

[0038] Example 2 Reference Figure 2 and Figure 3 Embodiment 2 of the present invention provides an electricity cost prediction method, which includes: like Figure 2 The dataset used is divided into three business scenarios: settlement-based meter reading and self-payment, which are considered periodic behaviors. Non-settlement-based meter reading, installment payments, and smart deductions are considered non-periodic behaviors. Simultaneously, the corresponding bill is issued, the corresponding user payment is processed, and the corresponding funds are credited to the account. In non-settlement-based meter reading, funds flow through three stages: issuance, payment, and receipt. In the installment payment and smart deduction scenarios, funds only flow through the payment and receipt stages. The prediction objective is to predict the daily cash flow for the following month based on historical data.

[0039] ① Periodic scenario prediction (meter reading during settlement, self-service payment) We constructed a Prophet model to predict daily cash flow for sub-scenarios. Daily cash flow data from each sub-scenarios was used as input to the training set, ensuring the model could learn over a time span of at least one year. Specifically, we customized the monthly periodicity of Prophet and added major holidays to the model, which served as external regression factors for prediction. Furthermore, in the settlement sub-business scenario, we added two exogenous variables, assigning a value of 1 on the 3rd and 4th of each month, and keeping them at 0 on other days, to strengthen the model's recognition and learning of amounts received on these two specific dates. Similarly, in the self-payment business scenario, we introduced two exogenous variables, assigning a value of 1 from the 13th to the 17th and from the 21st to the 27th of each month, and keeping them at 0 on other days, to further enhance the model's recognition and learning of amounts received during these time periods.

[0040] ② Non-periodic scenario prediction (non-settlement meter reading, smart deduction, and installment payment): Taking the construction of a model for predicting non-settlement meter reading scenarios as an example.

[0041] Estimating the daily issuance amount for non-settlement meter readings after day d: Calculate the average of the total issuance amount for non-settlement meter readings after day d in the previous three months, and use this as the estimated total issuance amount after day d. Further, calculate the specific proportion of the daily non-settlement meter reading amount to the total issuance amount after day d in each month from the end of day d in the past three months and the same period last year. Based on this, simulate the estimated daily issuance amount.

[0042] Key parameters are defined as follows: The amount of non-settlement meter readings issued but not yet paid before day d is obtained by subtracting the amount paid for non-settlement meter readings from the amount already issued for non-settlement meter readings. This amount is then uniformly considered as being issued on day d. Simultaneously, the "Daily Non-Settlement Meter Reading Collection Fee = Estimated Daily Non-Settlement Meter Reading Issuance Fee * (1 - Non-Settlement Issuance Pre-Collection Ratio) * Non-Settlement Meter Reading Receivable Recovery Rate" is defined. In the sub-models for the six user categories, the predicted parameters "Non-Settlement Issuance Pre-Collection Ratio" and "Non-Settlement Receivable Recovery Rate" both use a three-period moving average of historical data.

[0043] Estimating the daily payment amount for non-settlement meter readings after day d: The model uses a weight vector to quantify the pattern from issuance to payment. Taking the estimated issuance amount for day k as an example, we explain its underlying mechanism. Specifically, the model allocates the payment ratio for the estimated issuance amount on day k. First, the payment ratio on the first day after issuance is set as w1 of the issuance amount on day k. Second, the model estimates that on the fifth working day after issuance, the payment ratio should be w2 of the issuance amount on day k. In addition to these two key payment points, the model also considers the payment ratio allocation for other days during the period, as shown in the following figures. Figure 3 As shown.

[0044] The total percentage of this payment is set as (1-w1-w2) of the amount issued on day k. This yields an estimated daily payment amount for non-settlement meter readings. The estimated daily amount received for non-settlement meter readings after day d is calculated by categorizing payment methods into four main types: same-day payment, next-day payment, first business day payment, and within seven days. Based on historical payment and receipt information for non-settlement meter readings, the pattern of the "payment-receipt" delay is statistically analyzed by comparing the number of days after payment and whether the corresponding date falls on a holiday. Combined with the real-time updated "payment-receipt" pattern, a rolling forecast of the daily cash flow for a specific type of user's non-settlement meter reading segment is ultimately achieved.

[0045] ③ Multi-scenario result aggregation The predicted values ​​of each sub-scenario are superimposed and scaled to output the total daily cash flow.

[0046] See details Figure 3To verify the effectiveness of the method in this embodiment, electricity sales data from a prefecture-level city in China's Southern Power Grid in 2023 were selected, covering three scenarios: residential areas (strong cyclicality), industrial parks (weak cyclicality), and commercial complexes (hybrid). The hardware platform used Inspur NF5280 servers (dual-socket Xeon Gold 6330, 512GB memory), with a data span of 36 months.

[0047] First, a periodic determination of electricity consumption scenarios is performed. A 30-day observation window is set to count the number of manual interventions (such as manual load correction). When the number of monthly interventions in the industrial park is less than 5, periodic detection begins. The daily load curve is segmented into weekly segments, and an improved sliding window is used to analyze the periodic correlation, with a threshold set at 0.8.

[0048] For residential areas with strong periodicity, a time-series model is used to decompose the trend, seasonal, and special event components. External variables include: the extreme high-temperature warning period in July 2023 and the red alert days for cold waves in December 2023. Weekly cycle parameters are adjusted using dynamic Fourier series, with the base frequency set to 7 days to adapt to weekly load fluctuations.

[0049] For industrial parks with weak cyclicality, a delay rule base for "production planning → electricity consumption declaration → actual load" is constructed. Three weights are generated based on Q3 2023 data: planned execution ratio (normal production days), temporary production adjustment ratio (order fluctuations), and equipment maintenance ratio (downtime days). Time decay coefficients are weighted at 0.6:0.25:0.15 based on data from the past three months.

[0050] When the forecast period includes statutory holidays, the business compensation mechanism is activated: identify load fluctuations within 5 days before and after National Day, and allocate 65% of the deviation to the adjusted workdays according to business type. For multinational corporations, the work calendar of the parent company's country is loaded; for example, German companies automatically match Easter shutdown plans.

[0051] Model parameters are updated every 3 days, and manual verification is triggered when the prediction deviation exceeds 15%. The control group uses an LSTM neural network and a SARIMA model.

[0052] Table 1: Comparison of Scene Judgment Accuracy (Sample Size: 182 Power Distribution Zones)

[0053] Table 2: Load Forecasting Error in Residential Areas (Unit: %)

[0054] Table 3: Industrial Park Delay Response Prediction (Unit: MW)

[0055] Table 4: Holiday Compensation Effects of Commercial Complexes

[0056] Table 5: Calendar Adaptation Effect of Multinational Enterprises

[0057] Table 6: Comparison of Model Computational Performance

[0058] Table 1 demonstrates the revolutionary nature of the scenario determination mechanism: the misjudgment rate of this invention is reduced to 5.5% in special scenarios such as hospitals, which is 84% ​​lower than that of traditional methods. The core breakthrough lies in the accurate capture of the composite pattern of "daytime load surge + nighttime baseline rise" during the emergency department expansion period by covariance analysis, while traditional methods misjudge 89% of special events as abnormal due to a single threshold.

[0059] Table 2-3 validates the creativity of multi-stage modeling: The forecast error for residential areas during the Spring Festival was 6.8% (Table 2), stemming from the holiday item injection mechanism. On days with extreme high temperatures, external event coupling kept the error to 5.3%, while the LSTM forecast error reached 23.6% due to the lack of weather warning fusion.

[0060] The prediction bias for emergency industrial order response is only 5.6MW (Table 3), which is 7 times more accurate than traditional models. The key lies in the dynamic rule base generating temporary adjustment weight of 0.78, which accurately reflects the three-shift production mode.

[0061] Table 4-5 illustrates the robust innovation of the system: The peak load forecast for Christmas Eve was 3.5MW (Table 4). Through the work schedule adjustment compensation mechanism, 65% of the increased load on entertainment venues was accurately allocated to the 20:00-22:00 period. The traditional solution, which ignores the cross-day delay, resulted in a forecast error of 26.4% for the closing time.

[0062] German companies achieved a 94.7% accuracy rate in Easter load forecasting (Table 5) due to an extended multinational calendar loading mechanism that automatically matched shutdown plans. Traditional methods, failing to adapt to local holidays, misclassified 63.7% of actual load reductions as failures.

[0063] Table 6 reveals the engineering performance advantages: historical data dependency is only 3 months, reducing storage requirements by 75%. This is attributed to the rolling update mechanism, where each incremental processing only requires recent data, while LSTM requires the full 24 months of data for training.

[0064] Example 3 This embodiment provides an electronic device, including: one or more processors; and a storage device storing one or more programs thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any of the implementations of Embodiment 1 above.

[0065] Example 4 This embodiment provides a computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, it implements the method described in any of the implementations of Embodiment 1 above.

[0066] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. The terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, unless otherwise explicitly specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting periodic settlement data that integrates nonlinear dynamic lag modeling, characterized in that, Includes the following steps: The number of manual operations on business data per unit time in the target business scenario is used as the intervention frequency. The intervention frequency is compared with a preset threshold. When the intervention frequency is lower than the preset threshold, a monthly periodic judgment is made to determine whether the business scenario is a periodic scenario or a non-periodic scenario. The settlement behavior of the target business scenario is broken down into the triggering stage, the processing stage, and the completion stage, and correspondingly associated with the historical triggering time series data, historical processing time series data, and historical completion time series data in the business data. For periodic scenarios, a time series prediction model is used to predict the completed time series data, and the prediction results of the completed time series data are calibrated by scaling transformation based on historical data from the same period. For non-periodic scenarios, the weights from the trigger stage to the processing stage and from the processing stage to the completion stage are dynamically allocated based on historical business data. The parameters are updated using the moving average method, and the processing delay is dynamically adjusted by incorporating holidays and work schedule adjustments, in order to generate prediction results for the completion time series data. Output the prediction results of the completed time series data, and dynamically update the model parameters and prediction values ​​as new data arrives.

2. The method for predicting periodic settlement data by incorporating nonlinear dynamic lag modeling as described in claim 1, characterized in that, The steps for performing monthly periodicity determination include: Let the completion time series be (x1, x2, ..., x...). n ), k=[n / 30], then let X be... i = (x i+1 x i+2， ...,x i+30 (), where k is the monthly coefficient, and n represents the nth day of the month. The monthly cycle correlation coefficient is defined as: ； Calculate the monthly subsequence correlation of historical completed sequence data for the target business scenario. When the correlation coefficient exceeds a preset threshold, monthly periodicity is confirmed and the scenario is determined to be periodic; otherwise, it is a non-periodic scenario.

3. The method for predicting periodic settlement data by incorporating nonlinear dynamic lag modeling as described in claim 1, characterized in that, The process of breaking down the target settlement behavior into a triggering phase, a processing phase, and a completion phase includes: The triggering phase corresponds to a business initiation event, i.e., a bill generation operation; the processing phase corresponds to a user response event, i.e., a fund transfer operation; and the completion phase corresponds to a fund settlement event, i.e., a settlement completion operation.

4. The method for predicting periodic settlement data by incorporating nonlinear dynamic lag modeling as described in claim 3, characterized in that, The time series forecasting model makes predictions by decomposing trend, seasonal, and holiday components. The time series forecasting model is expressed by the following formula: Where g(t) is the trend component of the data, h(t) is the holiday component model, s(t) is the seasonal component model, and εt is the exogenous variable; Exogenous variables are introduced as inputs to the time series forecasting model, including policy adjustment points and system upgrade time points; The seasonal model is constructed using Fourier series and dynamically adjusts the period length to adapt to the monthly fluctuation patterns of different business scenarios.

5. The method for predicting periodic settlement data by incorporating nonlinear dynamic lag modeling as described in claim 1, characterized in that, The specific steps for completing the prediction results of time series data through scaling calibration based on historical contemporaneous data are as follows: Scale the time series data predicted by the trained time series model for the next month: ； Among them, W′ K ={W′ 1，K , W′ 2，K , ..., W′ tk，K } and W′ K+1 ={W′ 1，K+1 , W′ 2，K+1 , ..., W′ tk+1 , K+1 } represent the completed time series data for month K and month K+1 of last year, respectively. K ={W 1，K W 2，K , ..., W tk，K } represents the completed time series data for month K of this year, Y K ={Y 1，K Y 2，K , ..., Y tk，K } represents the predicted sequence obtained in the previous step, Y i,K The data is based on historical data and scaled, where j is the j-th day of the K-th month.

6. The method for predicting periodic settlement data by incorporating nonlinear dynamic lag modeling as described in claim 1, characterized in that, The modeling of non-periodic behavior includes: quantifying the weight allocation patterns of the triggering stage, processing stage, and completion stage based on historical data, and updating parameters using moving averages; The triggering phase prediction includes: Calculate the moving average based on the sum of the monthly trigger sequences for the previous three months; Based on the daily distribution pattern of the trigger sequence in the same period last year, generate the daily trigger volume allocation weight for the next month; The estimated value of the trigger time series data is obtained by multiplying the moving average value by the trigger amount weight; The processing stage prediction includes: Set the processing weight for the current day, the processing weight for delayed working days, and the weight allocated over a period; The triggered time series data forecast values ​​are allocated to the corresponding processing days according to three weight categories to obtain the processed time series forecast values, where: The weighting of processing on the current day will be allocated to the next day. The weighting of delayed working days is allocated to the 5th working day; The remaining value is evenly distributed among the working days during the period; The completion stage prediction includes: Three types of delay weights are calculated based on historical data from the previous three months: weight for completion on the same day, weight for completion on the next day, and weight for delay due to holidays. The processing time series estimates are allocated to the corresponding completion dates according to lag weights to obtain the completion time series estimates, where: Weight allocation is completed and carried over to the same day; The weight allocation will be completed the following day and carried over to the next day. The weighting of holidays will be postponed to the first non-holiday period.

7. The method for predicting periodic settlement data by incorporating nonlinear dynamic lag modeling as described in claim 6, characterized in that, The dynamic updating of model parameters and predicted values ​​when new data arrives includes: When actual business data arrives, the dynamic weight allocation parameters are recalculated; Update the moving average using a rolling time window; Perform business rule verification on the prediction results, and trigger manual review when the deviation exceeds the tolerance threshold.

8. A method for predicting electricity costs, characterized in that, Includes the following steps: Collect historical electricity bill datasets; Based on the periodic settlement data prediction method that integrates nonlinear dynamic lag modeling as described in any one of claims 1-7, multiple business scenarios in the historical electricity bill dataset are divided into periodic scenarios and non-periodic scenarios, and the daily electricity bill prediction data for the following month is obtained.

9. An electronic device, characterized in that... include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A computer-readable storage medium having executable instructions stored thereon, characterized in that... When executed by the processor, this instruction causes the processor to implement the method of any one of claims 1-7.

Images