Power grid material purchasing demand generation method and device for uncertain scene

Abstract

The invention discloses a power grid material purchasing demand generation method and device for an uncertain scene. The method comprises the steps of obtaining power grid material feature data of a to-be-generated purchasing amount; on the basis of a pre-constructed interval prediction model, demand prediction is carried out on the power grid material feature data, and an initial demand prediction interval in a target period is given; determining a historical prediction residual error of the interval prediction model, correcting the initial demand prediction interval, and giving a calibration demand prediction interval; and based on a pre-constructed purchase decision model, the calibration demand prediction interval is analyzed and processed, and the optimal purchase quantity in the target period is given. According to the method, the demand uncertainty can be accurately captured, and the consistency of the purchase quantity and the actual demand is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power grid material management, and particularly relates to a power grid material procurement demand generation method and device for uncertain scenarios. BACKGROUND

[0002] The power grid material supply chain is the core support for ensuring the safe and stable operation of the power system. At present, with the accelerated construction of new power systems, the expansion of new energy grid connection, and the continuous promotion of rural power grid transformation projects, the demand for power grid materials presents significant complexity and uncertainty. The demand is not only affected by traditional factors such as seasonal maintenance, infrastructure planning, and extreme weather, but also by structural changes caused by the access of new types of loads such as distributed photovoltaic and electric vehicle charging piles, resulting in further increased demand fluctuations. This high degree of uncertainty poses a serious challenge to material inventory management, with grassroots warehouses generally facing the dilemma of "urgent need for no goods or inventory accumulation". In some areas, the long delivery period of spare parts and the mismatch of models lead to delay in repair, while inventory accumulation results in a large amount of capital occupation, and the whole life cycle cost control pressure is prominent.

[0003] Traditional power grid material inventory management relies on experience quota or simple prediction model, which has obvious shortcomings: first, the single-point prediction method is used in the prediction link, the asymmetricity of demand distribution and the influence of prediction residual are not fully considered, which leads to large deviation between the prediction interval and the actual demand; second, the optimization process is not tolerant enough to uncertainty, and it is difficult to cope with demand fluctuations caused by multiple factors, which easily leads to the coexistence of "surplus and shortage". For example, patent CN106600032A discloses a kind of inventory material demand prediction method and device, the method comprises: according to initial investment plan, the planned investment of each project and the engineering property corresponding to each project are obtained;According to the engineering property corresponding to each project, the material small class contained in each project is obtained, the corresponding relationship list of the preset material small class and the engineering property and the prediction model is found, and the prediction model corresponding to each material small class is obtained;The planned investment is input into the prediction model corresponding to the corresponding material small class, and the predicted demand of each material small class in each project is calculated;The predicted demand of each material small class in different projects is added, and the total predicted demand of each material small class is calculated. For another example, patent CN118428862A discloses a kind of inventory material procurement prediction method and system, the method comprises: identifying abnormal data and filling missing values on historical data;Generate an initial predicted material proportion table, adjust the initial predicted material proportion table;Generate a number of predicted procurement lists;Respectively, the first prediction algorithm and the second prediction algorithm are used to predict the procurement materials based on the number of predicted procurement lists, and the initial prediction value is generated based on the first prediction value and the second prediction value;Based on the number of initial prediction values and the real value of procurement materials at the current time, the predicted procurement list is iteratively updated and adjusted until the iteration condition is met, and the updated predicted procurement list is output, and the final prediction value is output based on the updated predicted procurement list.

[0004] Therefore, there is an urgent need for a procurement demand generation method that can accurately capture demand uncertainty and balance risk and cost. SUMMARY

[0005] In view of the defects in the prior art, the present application provides a power grid material procurement demand generation method and device in an uncertain scenario. The demand uncertainty can be accurately captured, and the consistency of procurement quantity and actual demand can be improved.

[0006] In a first aspect, the present application provides a power grid material procurement demand generation method in an uncertain scenario, comprising: obtaining power grid material feature data for which procurement quantity is to be generated; based on a pre-constructed interval prediction model, predicting demand based on the power grid material feature data, and giving an initial demand prediction interval in a target period; The historical prediction residual of the interval prediction model is determined, the initial demand prediction interval is corrected, and a calibrated demand prediction interval is given. Based on the pre-constructed procurement decision model, the calibrated demand prediction interval is analyzed and processed, and the optimal procurement quantity in the target period is given.

[0007] Further, before the demand prediction of the power grid material characteristic data, further comprising: The power grid material characteristic data is preprocessed, and the preprocessed power grid material characteristic data is given.

[0008] Further, the pre-construction of the interval prediction model comprises: The historical power grid material characteristic data set and the corresponding historical demand data are obtained, and the training set is determined; The upper confidence level and the lower confidence level are determined, and the corresponding upper prediction model and lower prediction model are constructed; Based on the training set, the parameters of the upper prediction model and the lower prediction model are trained, and the model parameters of the upper prediction model and the lower prediction model are determined by minimizing the sum of the weighted absolute value deviation of the historical demand data and the demand prediction data as the target of the optimization function.

[0009] Further, the demand prediction of the power grid material characteristic data is performed to give an initial demand prediction interval in a target period, comprising: The upper prediction model and the lower prediction model of the model parameters are used to predict the demand of the power grid material characteristic data to be generated, and the upper limit value and the lower limit value are given; Based on the upper limit value and the lower limit value, the initial demand prediction interval is determined.

[0010] Further, the historical prediction residual of the interval prediction model is determined, the initial demand prediction interval is corrected, and a calibrated demand prediction interval is given, comprising: Another historical power grid material characteristic data set and corresponding historical demand data are obtained to form a validation set; Based on the interval prediction model, the historical power grid material characteristic data set of the validation set is predicted, and the historical demand data of the validation set is combined to determine the quantile prediction residual; Based on the quantile prediction residual and combined with the preset scale factor, the initial demand prediction interval is asymmetrically calibrated to determine the calibrated demand prediction interval.

[0011] Further, based on the interval prediction model, the historical power grid material characteristic data set of the validation set is predicted, and the historical demand data of the validation set is combined to determine the quantile prediction residual, comprising: Based on the upper bound prediction model and the lower bound prediction model, the historical power grid material feature data set of the verification set is demand predicted, and the upper bound prediction demand and the lower bound prediction demand are given. Based on the average of the upper bound prediction demand and the lower bound prediction demand, and combined with the historical demand data of the verification set, the corresponding quantile prediction residual is determined.

[0012] Further, based on the quantile prediction residual and combined with the preset scale factor, the initial demand prediction interval is asymmetrically calibrated to determine the calibrated demand prediction interval, including: Based on all quantile prediction residuals and corresponding mean values, the residual standard deviation and the deviation coefficient are determined. Based on the preset scale factor and the deviation coefficient, the upper bound skewness calibration coefficient and the lower bound skewness calibration coefficient are determined. Based on the residual standard deviation, the upper bound and the lower bound of the initial demand prediction interval are modified respectively combined with the upper bound skewness calibration coefficient and the lower bound skewness calibration coefficient to determine the calibrated demand prediction interval.

[0013] Further, the pre-construction of the procurement decision model includes: The historical demand data corresponding to the historical power grid material feature data set is obtained. Based on the historical demand data, the empirical distribution and the fuzzy set radius are determined, and the demand fuzzy set is given. Based on the demand fuzzy set, the shortage penalty cost and the material holding cost are determined, and the total cost is given. In the fuzzy set, the optimization function with the objective of minimizing the upper bound of the total cost is determined, and the procurement decision model is given.

[0014] Further, the calibrated demand prediction interval is analyzed and processed to give the optimal procurement quantity in the target period, including: The calibrated demand prediction interval is taken as the constraint condition of the procurement decision model, and the procurement decision model is strongly dual transformed; The strongly dual transformed procurement decision model is solved by a convex optimization solver or a linear programming solver to give the optimal procurement quantity.

[0015] In a second aspect, the present application also provides an uncertain scenario power grid material procurement demand generation device, which adopts the above uncertain scenario power grid material procurement demand generation method, and the device includes: A material data acquisition module is configured to acquire power grid material feature data for which procurement quantity is to be generated. A demand interval determination module is configured to determine an initial demand prediction interval in a target period based on a pre-constructed interval prediction model and perform demand prediction on the power grid material feature data. a demand interval optimization module, configured to determine a historical prediction residual of the interval prediction model, correct an initial demand prediction interval, and give a calibrated demand prediction interval; a purchase quantity determination module, configured to analyze and process the calibrated demand prediction interval based on a pre-constructed purchase decision model, and give an optimal purchase quantity in a target period.

[0016] Further, the device comprises a prediction model construction module, configured to: obtain a set of historical power grid material feature data and corresponding historical demand data, and determine a training set; determine an upper confidence level and a lower confidence level, and construct a corresponding upper limit prediction model and a lower limit prediction model; train the upper limit prediction model and the lower limit prediction model based on the training set, and determine model parameters of the upper limit prediction model and the lower limit prediction model through an optimization function with the goal of minimizing the sum of weighted absolute value deviations between the historical demand data and demand prediction data.

[0017] Further, the demand interval determination module is further configured to: perform demand prediction on power grid material feature data for which a purchase quantity is to be generated through the upper limit prediction model and the lower limit prediction model with determined model parameters, and give an upper limit value and a lower limit value; determine an initial demand prediction interval based on the upper limit value and the lower limit value.

[0018] Further, the demand interval optimization module is configured to: obtain another set of historical power grid material feature data and corresponding historical demand data to form a validation set; perform demand prediction on the set of historical power grid material feature data of the validation set based on the interval prediction model, and determine quantile prediction residuals in combination with the historical demand data of the validation set; perform asymmetric calibration on the initial demand prediction interval based on the quantile prediction residuals and a pre-set scale factor, and determine a calibrated demand prediction interval.

[0019] Further, the demand interval optimization module is configured to: perform demand prediction on the set of historical power grid material feature data of the validation set based on the upper limit prediction model and the lower limit prediction model, and give an upper limit prediction demand and a lower limit prediction demand; determine corresponding quantile prediction residuals based on an average value of the upper limit prediction demand and the lower limit prediction demand, and in combination with the historical demand data of the validation set.

[0020] Further, the demand interval optimization module is further configured to: determine a residual standard deviation and a deviation coefficient based on all quantile prediction residuals and corresponding average values; determine the upper bound skewness calibration coefficient and the lower bound skewness calibration coefficient based on the preset scale factor and the deviation coefficient; based on the residual standard deviation, in combination with the upper bound skewness calibration coefficient and the lower bound skewness calibration coefficient respectively, correct the upper bound value and the lower bound value of the initial demand prediction interval, and determine the calibrated demand prediction interval.

[0021] Further, the device further comprises a decision model construction module, which is used for: acquiring historical demand data corresponding to the historical power grid material feature data set; based on the historical demand data, determine the empirical distribution and the fuzzy set radius, and give the demand fuzzy set; based on the demand fuzzy set, determine the shortage penalty cost and the material holding cost, and give the total cost; within the fuzzy set, determine the optimization function with the minimum upper bound of the total cost as the target, and give the procurement decision model.

[0022] Further, the procurement quantity determination module is also used for: taking the calibrated demand prediction interval as a constraint condition of the procurement decision model, and performing strong duality transformation on the procurement decision model; by a convex optimization solver or a linear programming solver, solve the strong duality transformed procurement decision model, and give the optimal procurement quantity.

[0023] The method provided by the present application for generating the procurement demand of power grid materials in an uncertain scenario at least has the following beneficial effects: (1) The initial demand prediction interval is determined by the interval prediction model, which can objectively quantify the uncertainty. By correcting the possible systematic defects of the initial demand prediction interval, the calibrated demand prediction interval is determined, and the uncertainty quantification with higher information quality is realized. The corrected and high-quality uncertainty interval (calibrated demand prediction interval) is analyzed and processed by the procurement decision model, the uncertainty information (including the calibrated demand prediction interval) is included in the decision framework of the procurement decision model, and the procurement quantity determination considering the high complexity and uncertainty of the power grid material demand is realized. The total cost of the inventory system can be significantly reduced while ensuring high supply reliability, effectively solving the dilemma of coexistence of "shortage" and "overstock" faced by the grassroots warehouse.

[0024] (2) The demand distribution is determined by the interval prediction model, and the initial demand prediction interval with clear probability significance is generated, which can not preset the error distribution, thereby having high robustness to abnormal values and being able to adaptively reflect the heteroscedasticity of demand fluctuation with feature changes. The procurement decision model can provide a basic probabilistic uncertainty description, improve the single-point prediction to interval prediction, and realize the quantification and information enhancement of demand uncertainty.

[0025] (3) By statistically analyzing the historical prediction residuals and determining the corresponding skewness coefficients, the systematic bias in the direction of uncertainty can be identified and quantified. Then, according to the skewness direction and amplitude, the initial demand prediction interval is asymmetrically stretched and calibrated to generate the calibrated demand prediction interval. The systematic bias introduced due to the inconsistency of error distribution can be corrected, so that the calibrated demand prediction interval can more truly reflect the inequality of demand uplink risk and downlink risk, and provide more accurate support set information for the robust optimization of the procurement decision model.

[0026] (4) The procurement decision model describes the distribution uncertainty by the determined fuzzy set radius and the fuzzy set centered on the empirical distribution, and takes the calibrated demand prediction interval as the support set constraint of the fuzzy set. The true distribution can be required to "approach" the historical experience through the fuzzy set, thereby avoiding strong distribution assumptions. In addition, the calibrated demand prediction interval constraint can anchor the uncertainty range in a reasonable region driven by data, avoiding excessive conservatism. In addition, through strong duality transformation, it can ensure that the procurement decision model can be efficiently and accurately solved in practical application. The results not only output the optimal procurement quantity, but also obtain the worst-case expected cost upper bound and the dual variable, so that the theoretically rigorous but computationally complex model can be transformed into an automated decision-making tool that can be embedded into the actual management system, and finally generate the procurement quantity corresponding instruction with the optimal cost in the worst case under the given uncertainty level and risk tolerance. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A curve diagram for the existing agent reward linear demand replenishment problem; Figure 2 A flowchart of a power grid material procurement demand generation method in an uncertain scenario provided by an embodiment of the present application; Figure 3 An architectural diagram of a power grid material procurement demand generation method provided by an embodiment of the present application; Figure 4 A flowchart of constructing an interval prediction model provided by an embodiment of the present application; Figure 5 A flowchart of giving a calibrated demand prediction interval provided by an embodiment of the present application; Figure 6 A flowchart of performing asymmetric calibration provided by an embodiment of the present application; Figure 7 A flowchart of constructing a procurement decision model provided by an embodiment of the present application; Figure 8 A schematic diagram of a power grid material procurement demand generation device in an uncertain scenario provided by the present application. DETAILED DESCRIPTION

[0028] In order to make the above technical solutions better understood, the above technical solutions will be described in detail below in combination with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0029] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. "Multiple" generally includes at least two.

[0030] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that the product or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such product or device. Without more limitations, the element defined by the sentence "including a" does not exclude the presence of additional identical elements in the product or device including the element.

[0031] The core of the present application is based on the "predict-then-optimize" framework, and improvements are made for its shortcomings. In data-driven decision-making, decision-makers usually first estimate unknown parameters from historical data (prediction), and then use these estimates as input optimization models to generate final decisions (optimization) for the best optimization problem. This process is called the Predict-Then-Optimize (PTO) framework. In order to formalize the optimization problem, consider a decision maker who maximizes the reward function F(z, θ) by choosing a decision z, where θ is an unknown parameter. By calling F(z, θ) a reward function, no generality is lost. If θ is known, the optimal decision will be z(θ)=arg min F(z, θ). In the PTO framework, the decision maker first uses historical data to produce an estimate , which is the "prediction" step; then the decision maker uses this estimate to solve the optimization problem, producing the decision , which is the "optimization" step. The final reward is . Since the function form of F and the mapping are known, the performance of PTO can be analyzed by studying the agent reward function, defined as .

[0032] PTO framework is popular for its simplicity and modularity. By treating prediction and optimization as separate steps, it allows mature statistical and machine learning methods to be leveraged for the prediction task. However, even if the parameter estimates produced by the prediction step are statistically unbiased, the final decision outcome can not be optimal. The root cause is that the performance of the optimization decision (i.e., the "agent reward function") is asymmetrically sensitive to parameter estimation errors. The following is an existing example: Take data-driven inventory replenishment with linear demand using PTO as an example: Set the revenue level p to maximize the revenue F(p, θ) = pd(p), where the demand function is d(p) = a - θp, and assume that a is known and the demand sensitivity θ is unknown. The optimal pricing p(θ) = a / (2θ). If the estimated value is used to set the revenue level , then the actual revenue (the agent reward function) is:

[0033] As shown in Figure 1 , the agent reward function reaches a maximum at the true value , but the shape is asymmetric, and the penalty of underestimating θ (i.e., θ ) is much greater than the penalty of overestimating θ by the same amount. Therefore, even for an unbiased estimator (satisfying ), the expected reward may still be systematically affected due to the significantly higher risk of underestimation.

[0034] Based on this, as shown in Figure 2 and Figure 3 , the present application provides a power grid material procurement demand generation method in an uncertain scenario, comprising: obtaining the characteristic data of the power grid material whose procurement quantity is to be generated; based on a pre-constructed interval prediction model, predicting the demand of the power grid material characteristic data, and giving an initial demand prediction interval in a target period; determining the historical prediction residual of the interval prediction model, correcting the initial demand prediction interval, and giving a calibrated demand prediction interval; based on a pre-constructed procurement decision model, analyzing and processing the calibrated demand prediction interval, and giving the optimal procurement quantity in the target period.

[0035] The present application further comprises, before predicting the demand of the power grid material characteristic data: The power grid material feature data is preprocessed, and the preprocessed power grid material feature data is given. The power grid material feature data can include deterministic features and predictable features. The deterministic features are known data, which can include time features, planned features, policy features, and historical statistical features (such as average demand in a predetermined period, historical same-period demand, etc.). The predictable features can be predicted by other known models, such as weather features predicted by a weather forecast model, load features predicted by a power load prediction model, and economic features predicted by a macroeconomic prediction model, construction progress features predicted based on engineering plans combined with historical progress rules, etc. Although the predictable features themselves also have uncertainties, the uncertainties of the predictable features are considered and included when the distribution robust optimization model (procurement decision model) is analyzed and processed. The preprocessing can include missing value processing, outlier processing, and feature standardization, which are all conventional data preprocessing schemes and will not be described here.

[0036] Before demand prediction of the power grid material feature data, a pre-constructed interval prediction model needs to be determined. Specifically, as shown in Figure 4 , the pre-construction of the interval prediction model can include: obtaining a set of historical power grid material feature data and corresponding historical demand data, and determining a training set; determining an upper confidence level and a lower confidence level, and respectively constructing a corresponding upper bound prediction model and a lower bound prediction model; training the upper bound prediction model and the lower bound prediction model based on the training set, and determining the model parameters of the upper bound prediction model and the lower bound prediction model through an optimization function with the minimum sum of weighted absolute value deviations between the historical demand data and the demand prediction data as the target.

[0037] After constructing the interval prediction model, demand prediction of the power grid material feature data can be performed to give an initial demand prediction interval in a target period, including: using the upper bound prediction model and the lower bound prediction model with determined model parameters to respectively predict the demand of the power grid material feature data for which the procurement quantity is to be generated, and giving an upper bound value and a lower bound value; determining the initial demand prediction interval based on the upper bound value and the lower bound value.

[0038] In actual application scenarios, in order to accurately depict the characteristics of material demand at different probability levels, quantile regression (QR) is used to construct the interval prediction model. Specifically, given the historical power grid material feature data x t of a predetermined period t, the quantile regression aims to find a function q τ (x t), such that the sum of weighted absolute deviations (i.e., the pinball loss function) is minimized, subject to the following relationship:

[0039] wherein, is the pinball loss function, u= ξ t -q , q=q τ (x t ). By setting the confidence level τ=τ0 and τ=1-τ0 respectively, the corresponding lower bound prediction model and upper bound prediction model, the value of τ0 can be 0.1, 0.2, etc. The function q τ (x t ) structure of the lower bound prediction model and the upper bound prediction model can be the same, and the difference is only the model / function parameters, which are trained by the historical power grid material feature data x t of the predetermined period t, and the model parameters of the lower bound prediction model and the upper bound prediction model are given by optimizing the pinball loss function with different values of the confidence level. By determining the lower bound prediction model and the upper bound prediction model, and based on the actual historical power grid material feature data in the future period to be predicted, the lower bound and the upper bound of the demand in the corresponding period can be obtained.

[0040] Compared with the traditional method assuming that the error term is independently distributed, the present application has the following significant advantages: 1. Strong robustness, not sensitive to outliers, can effectively filter out accidental data noise and avoid single extreme value to deviate the overall prediction result; 2. Heteroscedastic adaptability, which can capture the dynamic range of demand fluctuation with feature variables. For example, during the peak of the project or in extreme weather, the demand fluctuation becomes larger, and the initial demand prediction interval generated by the quantile regression will automatically widen, thereby providing a more informative "possible space" for the decision model (procurement decision model). The initial demand prediction interval preliminarily delimits the feasible region of decision-making, which is the key interface connecting data-driven prediction and operational optimization decision-making.

[0041] Although the initial demand prediction interval has certain statistical significance, it often implies the assumption that the error distribution is symmetric or uniform. However, the prediction error of power grid materials demand often has significant skewness. For example, due to the existence of sudden repair tasks, the probability of actual demand surge (positive deviation) is often much higher than that of demand sudden decrease (negative deviation), i.e., the residual distribution presents right-skewed characteristics. If this skewness is ignored and the initial demand prediction interval is directly used, it will lead to underestimation of the upper tail risk, thereby causing stockout. Therefore, the present application adopts a posterior adjustment mechanism based on residual skewness.

[0042] Specifically, as Figure 5As shown, after determining the initial demand prediction interval, the historical prediction residual of the interval prediction model can be determined, the initial demand prediction interval is corrected, and the calibrated demand prediction interval is given, including: obtaining another set of historical power grid material feature data and corresponding historical demand data to form a verification set; based on the interval prediction model, the historical power grid material feature data set of the verification set is predicted, and the historical demand data of the verification set is combined to determine the quantile prediction residual; based on the quantile prediction residual and combined with the preset scale factor, the initial demand prediction interval is asymmetrically calibrated to determine the calibrated demand prediction interval.

[0043] wherein, based on the interval prediction model, the historical power grid material feature data set of the verification set is predicted, and the historical demand data of the verification set is combined to determine the quantile prediction residual, including: based on the upper bound prediction model and the lower bound prediction model, the historical power grid material feature data set of the verification set is predicted to give the upper bound prediction demand and the lower bound prediction demand; based on the average of the upper bound prediction demand and the lower bound prediction demand, and combined with the historical demand data of the verification set, the corresponding quantile prediction residual is determined.

[0044] As shown Figure 6 based on the quantile prediction residual and combined with the preset scale factor, the initial demand prediction interval is asymmetrically calibrated to determine the calibrated demand prediction interval, which can include: based on all quantile prediction residuals and corresponding mean values, the residual standard deviation and the deviation coefficient are determined; based on the preset scale factor and the deviation coefficient, the upper bound skewness calibration coefficient and the lower bound skewness calibration coefficient are determined; based on the residual standard deviation, the upper bound skewness calibration coefficient and the lower bound skewness calibration coefficient are combined to correct the upper bound value and the lower bound value of the initial demand prediction interval, and the calibrated demand prediction interval is determined. The upper bound value and the lower bound value correspond to the upper limit and the lower limit of the interval respectively.

[0045] In actual application scenarios, first, the quantile prediction residual on the verification set is calculated , ξ t is the actual historical demand data in the verification set, q' τ (x t ) is the average of the upper bound prediction demand and the lower bound prediction demand corresponding to the verification set, and the skewness coefficient of the overall residual is calculated . Specifically, it satisfies the following relationship:

[0046] According to the sign and size of kappa, a scale factor alpha and a residual standard deviation S are introduced e , define the skewness calibration coefficient:

[0047] In the formula, delta up is the upper limit adjustment amount, delta low is the lower limit adjustment amount; Based on this, the corrected calibration demand prediction interval , meet the following relationship:

[0048] Specifically, when the error is right biased (kappa>0), the upper limit of the prediction is automatically raised, and the safety stock is increased to resist the risk of shortage; when the error is left biased (kappa<0), the lower limit of the prediction is appropriately reduced to avoid blind stockpiling. Based on this correction, the calibration demand prediction interval can be obtained, and a procurement decision model (distribution robust optimization (DRO) model based on Wasserstein distance) can be constructed and solved. The procurement decision model introduces a fuzzy set Omega centered on the empirical distribution and with a Wasserstein distance epsilon as the radius, and the corrected calibration demand prediction interval is explicitly used as the support set constraint of the fuzzy set. Through strong duality theory, the infinite-dimensional optimization problem is transformed into a finite-dimensional convex optimization problem.

[0049] In material inventory management, demand prediction can provide important reference information for inventory planning, but relying only on prediction results is still insufficient to achieve scientific inventory decision. Existing prediction models usually give point prediction values and interval ranges of future demand, but existing prediction models can only "describe" the possible demand state in the future, and cannot directly give the optimal ordering strategy under different cost constraints. That is, the demand prediction result does not have decision significance, cannot dynamically balance between cost, risk and service level, and cannot guarantee the matching degree between inventory level and actual demand.

[0050] In the actual scenario of the power grid, the determination of the procurement quantity is a key link in the entire material supply chain management. Too much procurement quantity will lead to inventory accumulation, capital occupation and storage cost increase; too little procurement quantity may cause material shortage, affecting the construction progress and power supply guarantee. Therefore, how to scientifically determine the optimal procurement quantity under the condition of demand prediction uncertainty is the core decision problem of power material management. The core of the procurement decision model of the present application is to set a clear objective function, and obtain the optimal decision value that best fits the actual situation by solving the objective function. In the inventory problem, the objective function can comprehensively consider factors such as material holding cost, shortage penalty cost, procurement cost, etc., and through optimization, the optimal inventory level or replenishment quantity is obtained, so that the total cost of the system is minimized.

[0051] Correspondingly, as shown in Figure 7 The pre-construction of the procurement decision model can include: Obtaining historical demand data corresponding to the historical power grid material feature data set; Based on the historical demand data, determine the experience distribution and fuzzy set radius, and give the demand fuzzy set; Based on the demand fuzzy set, determine the shortage penalty cost and material holding cost, and give the total cost; In the fuzzy set, determine the optimization function with the objective of minimizing the upper bound of the total cost, and give the procurement decision model.

[0052] The analysis and processing of the calibrated demand prediction interval can include: Take the calibrated demand prediction interval as the constraint condition of the procurement decision model, and perform strong duality transformation on the procurement decision model; Solve the strong duality transformed procurement decision model through a convex optimization solver or a linear programming solver, and give the optimal procurement quantity.

[0053] Wherein, in the pre-construction of the procurement decision model, the historical power grid material feature data set used corresponds to the training set in the pre-construction of the interval prediction model, and the historical power grid material feature data sets of the validation set and the training set are different, which can be data in different historical periods.

[0054] In actual application scenarios, the construction and solving process of the procurement decision model specifically includes: When the demand distribution is known, the best order quantity is the b / (b+h) quantile of the demand distribution, where b is the unit shortage penalty cost and h is the unit inventory overstock cost. However, it is usually difficult to obtain accurate probability distribution information, and only partial incomplete knowledge of the distribution can be obtained, such as data samples, historical data, statistical information, etc. By constructing a fuzzy set containing all probability distributions, we hope to find an optimal strategy that can approach the optimal performance regardless of the changes in the true demand distribution. The initial objective function can be expressed as follows:

[0055] In the formula, is the expectation, ξ is a random variable representing the uncertain actual demand, P is the true probability distribution of the actual demand ξ, Ξ is the value space of the actual demand ξ, x is the decision variable representing the planned procurement (or inventory replenishment) quantity, i.e. the procurement quantity, (x-ξ) + represents the inventory remaining amount, (x-ξ) + =max(x-ξ,0), (ξ-x) + represents the shortage amount, (ξ-x) + =max(ξ-x,0) The distribution robust optimization is performed based on the initial objective function. Specifically, the construction of the fuzzy set in the distribution robust optimization is crucial to enhance the generalization ability and adaptability to uncertainty of the procurement decision model. The fuzzy set constructed by using the Wasserstein distance can reflect more accurate distribution differences and adaptively cope with the changing data environment, thereby improving the performance of the procurement decision model in uncertainty and disturbance. Moreover, the Wasserstein distance has a good mathematical theoretical basis, and its calculation process and optimization mode are clearly defined and explained in probability theory and transportation theory. Therefore, the fuzzy set constructed based on the Wasserstein distance is more rigorous in theory and is easy to obtain a clear optimization solution. In the Wasserstein distance, the fuzzy set is represented as a sphere, which refers to a set of distributions within a certain range centered on an empirical distribution . The optimization function of the procurement decision model within the fuzzy set and aiming to minimize the upper bound of the total cost can be represented as:

[0056] where P is an actual demand distribution, represents the constructed empirical distribution, γ is a transport scheme, is a joint distribution with the marginal distribution being the demand distribution and the empirical distribution, is used to calculate the Wasserstein distance, μ and v are two probability distributions to be compared, and correspond to the actual demand distribution and the empirical distribution, respectively, Γ(μ,v) is a set of all joint probability distributions γ with μ and v as the marginal distribution, d(ξ,ζ) is a distance metric between ξ and ζ, which can be the Euclidean distance, and dγ(ξ,ζ) is the probability differential of the joint distribution γ, represents the distance between the demand distribution and the empirical distribution, represents the spherical radius of the fuzzy set, which allows the maximum distance to deviate from the center empirical distribution ξ represents a random demand amount under the actual demand distribution P, and ζ represents a random amount under the empirical distribution . The value of p represents the order of the Wasserstein distance, which measures the difference between the demand distribution and the empirical distribution, especially the relative sensitivity between the demand distribution and the empirical distribution. A smaller order pays more attention to local differences, that is, smaller deviations between the demand distribution and the empirical distribution. When the distribution is uneven or there are extreme points, the Wasserstein distance is more sensitive to these uneven points. A larger order pays more attention to overall differences, especially changes in large deviations or extreme points, and has low sensitivity to extreme points.

[0057] The experience distribution is the core input of the whole model, reflecting the demand probability structure estimated from the limited sample, and is the benchmark distribution for constructing the fuzzy set. In actual scenarios, the true distribution of demand cannot be known, but historical demand data is often available, and the historical data can be used to construct the fuzzy set and optimize the constructed ambiguous fuzzy set. Specifically, the Dirac measure is a special probability measure that concentrates all the mass at a single point, i.e., taking the value 1 at a certain specific point and 0 at other points. It is commonly used in the construction of experience distribution in Wasserstein distance, satisfying the following relationship:

[0058] wherein, is the Dirac measure, is the support point on the Dirac measure, corresponding to a historical demand data in the training set, is the set on the definition domain, i.e., the set of all historical demand data corresponding to the training set.

[0059] The weighted sum of multiple Dirac measures can be used to construct the experience distribution, assuming and , is the i-th historical demand sample value, i.e., the historical demand data, wherein . The experience distribution constructed based on the training set can be expressed as follows:

[0060] represents the Dirac measure of the i-th historical demand sample value.

[0061] The experience distribution is essentially a discrete probability distribution, with all its probability mass concentrated on a finite number of observed sample points.

[0062] In the procurement decision model, the radius epsilon determines the size of the fuzzy set. When the radius is large, the distribution contained in the fuzzy set is more extensive, reflecting a higher tolerance for uncertainty in the model. Therefore, the model needs to be optimized under a wider distribution set to ensure that it performs well under all these distributions. The final decision result (procurement quantity) will be conservative, and the calculation will be more difficult. When the radius is infinite, the model degenerates into a robust optimization model. Conversely, a smaller radius limits the distribution changes during optimization, considering similar situations to the central distribution, with lower tolerance for changes. The optimization process is flexible, but it is not sensitive to unconsidered disturbances or distribution changes, leading to insufficient robustness. When the radius is 0, the model degenerates into a stochastic optimization model. Therefore, a reasonable choice of fuzzy set radius is crucial for the robustness and generalization ability of the distribution robust inventory optimization model. Therefore, the present application provides a calculation formula for the radius of the fuzzy set, satisfying the following relationship:

[0063] wherein, represents the confidence level of the fuzzy set, N represents the sample number, a0 is a Lagrange multiplier or a dual variable, and D is a constant, is the mean value of the historical demand sample (historical demand data).

[0064] Based on the empirical distribution constructed according to the Dirac measure, the procurement decision model can be expressed as:

[0065] In the formula, λ is the first dual variable introduced by the strong dual transformation, that is, the distribution uncertainty penalty coefficient. In the inventory optimization problem, the power grid project decision maker often expects that the engineering materials can be supplied on time when they are needed, and is willing to bear greater inventory management cost risk, so λ≥b>h is met; μ1 and μ2 are the second dual variable and the third dual variable introduced by the strong dual transformation, respectively, corresponding to the Lagrange multipliers of the lower bound constraint and the upper bound constraint of the calibrated demand prediction interval.

[0066] In order to transform the optimization function with the minimum total cost upper bound as the goal into a form convenient for solving, it is necessary to analyze and characterize the upper bound corresponding to the "worst case". In the procurement decision model, demand prediction is introduced as a constraint condition. Specifically, the demand prediction not only gives a specific predicted value, but also accompanies an interval with upper and lower bounds. Therefore, in addition to the fuzzy set constraint, there are the following probability constraints:

[0067] By combining demand prediction with the uncertainty interval, the robustness of the procurement decision model can be improved. By introducing the predicted value and its interval into the optimization goal, the additional information of the prediction result on future demand can be constrained while retaining the ability of the model to handle uncertainty. It can effectively alleviate the deviation caused by relying only on historical data, enhance the adaptability when facing actual demand changes, and thus improve the precision and robustness of the procurement decision model output.

[0068] Among them, the procurement decision model is in the fuzzy set, and the optimization function with the minimum total cost upper bound as the goal is determined. The corresponding model is simplified and expressed as:

[0069] In the formula, f(x, ξ) is a single-period inventory cost function, ; After strong dual transformation, the model can be expressed as:

[0070] In the formula, S i is an auxiliary variable for the i-th historical demand sample, and ξ represents an uncertain actual demand that may occur in the future.

[0071] The procurement decision model converted by strong duality can be used to make replenishment decisions and effectively improve the precision and robustness of the replenishment decision. Specifically, the solution of the procurement decision model converted by strong duality can be realized by a convex optimization solver (such as Gurobi, CVX, etc.) or a linear programming solver to determine the final replenishment decision (optimal procurement quantity).

[0072] The above technical solutions can at least include the following beneficial effects: (1) The initial demand prediction interval is determined by the interval prediction model, and the uncertainty can be objectively quantified. By correcting the possible systematic defects of the initial demand prediction interval, the calibrated demand prediction interval is determined, and the uncertainty quantification with higher information quality is realized. The corrected and high-quality uncertainty interval (calibrated demand prediction interval) is analyzed and processed by the procurement decision model, the uncertainty information (including the calibrated demand prediction interval) is included in the decision framework of the procurement decision model, and the procurement quantity determination considering the high complexity and uncertainty of power grid material demand is realized. The total cost of the inventory system can be significantly reduced while ensuring high supply reliability, effectively solving the dilemma of coexistence of "shortage" and "overstock" faced by the grassroots warehouse.

[0073] (2) The demand distribution is determined by the interval prediction model to generate an initial demand prediction interval with clear probability significance, which can not pre-set error distribution, thereby having high robustness to abnormal values and being able to adaptively reflect the heteroscedasticity of demand fluctuation with feature changes. The procurement decision model can provide basic probabilistic uncertainty characterization, upgrade single-point prediction to interval prediction, and realize quantification and information enhancement of demand uncertainty.

[0074] (3) By statistically analyzing the historical prediction residual and determining the corresponding skewness coefficient, the systematic deviation of the direction of uncertainty can be identified and quantified. Then, according to the skewness direction and amplitude, the initial demand prediction interval is calibrated by asymmetric stretching to generate the calibrated demand prediction interval. The systematic deviation introduced due to the inconsistency of the error distribution can be corrected, so that the calibrated demand prediction interval can more truly reflect the inequality of the uplink risk and downlink risk of demand, and provide more accurate support set information for the robust optimization of the procurement decision model.

[0075] (4) The procurement decision model describes distribution uncertainty by a fuzzy set with a determined radius and a fuzzy set centered on an empirical distribution, and takes the calibrated demand prediction interval as a support set constraint. The fuzzy set can require the true distribution to "approach" the historical experience, thereby avoiding strong distribution assumptions. In addition, the calibrated demand prediction interval constraint can anchor the uncertainty range in a reasonable region driven by data, avoiding excessive conservatism. In addition, through strong duality transformation, it can ensure that the procurement decision model can be efficiently and accurately solved in practical application. The results not only output the optimal procurement quantity, but also obtain the worst-case expected cost upper bound and the dual variable, so that the theoretically rigorous but computationally complex model can be transformed into an automated decision-making tool that can be embedded in a practical management system, and finally generate a procurement quantity corresponding instruction with the worst-case optimal cost under the given uncertainty level and risk tolerance.

[0076] As shown in Figure 8 The present application also provides a power grid material procurement demand generation device for uncertain scenarios, which adopts the above-mentioned power grid material procurement demand generation method for uncertain scenarios. The device comprises: a material data acquisition module for acquiring power grid material characteristic data for which procurement quantity is to be generated; a demand interval determination module for predicting demand based on a pre-constructed interval prediction model, and giving an initial demand prediction interval in a target period based on the power grid material characteristic data; a demand interval optimization module for determining historical prediction residuals of the interval prediction model, correcting the initial demand prediction interval, and giving a calibrated demand prediction interval; a procurement quantity determination module for analyzing and processing the calibrated demand prediction interval based on a pre-constructed procurement decision model, and giving an optimal procurement quantity in the target period.

[0077] Further, the device comprises a prediction model construction module for: acquiring a set of historical power grid material characteristic data and corresponding historical demand data, and determining a training set; determining an upper confidence level and a lower confidence level, and respectively constructing an upper limit prediction model and a lower limit prediction model; training parameters of the upper limit prediction model and the lower limit prediction model based on the training set, and determining model parameters of the upper limit prediction model and the lower limit prediction model through an optimization function with the objective of minimizing the sum of weighted absolute value deviations between historical demand data and demand prediction data.

[0078] Further, the demand interval determination module is further used for: predicting demand for the power grid material characteristic data for which procurement quantity is to be generated through the upper limit prediction model and the lower limit prediction model with determined model parameters, and giving an upper limit value and a lower limit value. determine an initial demand prediction interval based on the upper bound value and the lower bound value.

[0079] Further, the demand interval optimization module is configured to: obtain another set of historical power grid material feature data and corresponding historical demand data to form a verification set; based on the interval prediction model, perform demand prediction on the set of historical power grid material feature data of the verification set, and combine the historical demand data of the verification set to determine quantile prediction residuals; based on the quantile prediction residuals and in combination with a preset scale factor, perform asymmetric calibration on the initial demand prediction interval to determine a calibrated demand prediction interval.

[0080] Further, the demand interval optimization module is configured to: based on the upper bound prediction model and the lower bound prediction model, perform demand prediction on the set of historical power grid material feature data of the verification set to give upper bound prediction demand and lower bound prediction demand; based on the average of the upper bound prediction demand and the lower bound prediction demand, and in combination with the historical demand data of the verification set, determine corresponding quantile prediction residuals.

[0081] Further, the demand interval optimization module is further configured to: based on all quantile prediction residuals and corresponding average values, determine residual standard deviation and deviation coefficient; based on a preset scale factor and the deviation coefficient, determine an upper bound skewness calibration coefficient and a lower bound skewness calibration coefficient; based on the residual standard deviation, in combination with the upper bound skewness calibration coefficient and the lower bound skewness calibration coefficient respectively, correct the upper bound value and the lower bound value of the initial demand prediction interval to determine a calibrated demand prediction interval.

[0082] Further, the device further comprises a decision model construction module configured to: obtain historical demand data corresponding to the set of historical power grid material feature data; based on the historical demand data, determine an empirical distribution and a fuzzy set radius, and give a demand fuzzy set; based on the demand fuzzy set, determine a stockout penalty cost and a material holding cost, and give a total cost; within the fuzzy set, determine an optimization function with the objective of minimizing the upper bound of the total cost, and give a procurement decision model.

[0083] Further, the procurement quantity determination module is further configured to: take the calibrated demand prediction interval as a constraint condition of the procurement decision model, and perform strong duality transformation on the procurement decision model; Solve the purchase decision model after strong dual conversion by convex optimization solver or linear programming solver, and give the optimal purchase quantity.

[0084] While the preferred embodiments of the application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to encompass within their scope all possible variations and modifications of the preferred embodiments. It is apparent that those skilled in the art can, without departing from the spirit or scope of the application, make various changes and modifications of the application. Thus, the application is intended to encompass all such changes and modifications as fall within the scope of the claims, together with all equivalents thereof.

Claims

1. A method for generating power grid material procurement requirements in uncertain scenarios, characterized in that, include: Obtain the characteristic data of power grid materials to be procured; Based on a pre-built interval prediction model, demand forecasting is performed on power grid material characteristic data, and the initial demand forecast interval within the target period is given. Determine the historical forecast residuals of the interval forecast model, correct the initial demand forecast interval, and give the calibrated demand forecast interval. Based on a pre-built procurement decision model, the calibration demand forecast range is analyzed and processed to provide the optimal procurement quantity within the target period.

2. The method for generating power grid material procurement requirements as described in claim 1, characterized in that, Before performing demand forecasting based on power grid material characteristic data, the following steps are also included: The power grid material characteristic data is preprocessed, and the preprocessed power grid material characteristic data is provided.

3. The method for generating power grid material procurement requirements as described in claim 1, characterized in that, The pre-construction of the interval prediction model includes: Obtain a set of historical power grid material characteristic data and corresponding historical demand data to determine the training set; Determine the upper and lower confidence levels, and construct the corresponding upper and lower bound prediction models respectively; The parameters of the upper bound prediction model and the lower bound prediction model are trained based on the training set, and the model parameters of the upper bound prediction model and the lower bound prediction model are determined by an optimization function that aims to minimize the sum of the weighted absolute value deviations between historical demand data and demand forecast data.

4. The method for generating power grid material procurement requirements as described in claim 3, characterized in that, Demand forecasting is performed based on power grid material characteristic data, providing an initial demand forecast interval for the target period, including: By determining the upper and lower bound prediction models of the model parameters, demand forecasting is performed on the power grid material characteristic data to be procured, and the upper and lower bound values ​​are given. The initial demand forecast interval is determined based on the upper and lower bound values.

5. The method for generating power grid material procurement requirements as described in any one of claims 1-4, characterized in that, Determine the historical forecast residuals of the interval forecasting model, revise the initial demand forecasting interval, and provide a calibrated demand forecasting interval, including: Obtain additional historical power grid material characteristic data sets and corresponding historical demand data to form a validation set; Based on the interval prediction model, demand prediction is performed on the historical power grid material characteristic data set of the validation set, and the quantile prediction residuals are determined by combining the historical demand data of the validation set. Based on the quantile prediction residuals and combined with a preset scaling factor, the initial demand prediction interval is asymmetrically calibrated to determine the calibrated demand prediction interval.

6. The method for generating power grid material procurement requirements as described in claim 5, characterized in that, Based on the interval forecasting model, demand forecasting is performed on the historical power grid material characteristic data set of the validation set. Combined with the historical demand data of the validation set, the quantile forecast residuals are determined, including: Both the upper bound prediction model and the lower bound prediction model are used to predict demand based on the historical power grid material characteristic data set of the validation set, and the upper bound prediction demand and the lower bound prediction demand are given. Based on the average of the upper and lower bounds of the predicted demand, and in conjunction with historical demand data from the validation set, the corresponding quantile prediction residuals are determined.

7. The method for generating power grid material procurement requirements as described in claim 6, characterized in that, Based on the quantile prediction residuals and combined with a preset scaling factor, the initial demand prediction interval is asymmetrically calibrated to determine the calibrated demand prediction interval, including: Based on the predicted residuals of all quantiles and their corresponding means, determine the residual standard deviation and the coefficient of deviation; Based on the preset scale factor and deviation coefficient, the upper bound skewness calibration coefficient and the lower bound skewness calibration coefficient are determined; Based on the residual standard deviation, the upper and lower bounds of the initial demand forecast interval are corrected by combining the upper and lower bound skewness calibration coefficients, respectively, to determine the calibration demand forecast interval.

8. The method for generating power grid material procurement requirements as described in claim 5, characterized in that, The pre-construction of the procurement decision model includes: Obtain historical demand data corresponding to the historical power grid material characteristic data set; Based on historical demand data, determine the empirical distribution and fuzzy set radius, and provide the demand fuzzy set; Based on the fuzzy set of demand, determine the stockout penalty cost and material holding cost, and give the total cost; Within the fuzzy set, determine the optimization function with the objective of minimizing the upper bound of total cost, and give the procurement decision model.

9. The method for generating power grid material procurement requirements as described in claim 8, characterized in that, The calibration demand forecast range is analyzed and processed to provide the optimal procurement quantity within the target period, including: The calibration demand forecast interval is used as a constraint on the procurement decision model, and a strong dual transformation is performed on the procurement decision model. The procurement decision model after strong dual transformation is solved by using a convex optimization solver or a linear programming solver to give the optimal procurement quantity.

10. A device for generating power grid material procurement needs under uncertain scenarios, characterized in that, The apparatus for generating power grid material procurement requirements under uncertain scenarios as described in any one of claims 1-9 includes: The material data acquisition module is used to acquire the characteristic data of power grid materials to be procured. The demand range determination module is used to predict the demand based on the power grid material characteristic data based on the pre-built range prediction model, and give the initial demand prediction range within the target period. The demand interval optimization module is used to determine the historical forecast residuals of the interval forecast model, correct the initial demand forecast interval, and provide a calibrated demand forecast interval. The procurement quantity determination module is used to analyze and process the calibration demand forecast range based on a pre-built procurement decision model, and give the optimal procurement quantity within the target period.

Images