Multi-source covariate irrigation load short-period prediction method, device and medium

Abstract

The application relates to a multi-source covariate irrigation load short-period prediction method, a device and a medium, and relates to the fields of power system load prediction and intelligent analysis of agricultural irrigation energy. The application is to solve the problems that the existing irrigation load short-period prediction method has limited modeling capability for exogenous factors, cannot effectively utilize user difference information, and is prone to lag or drift in prediction. The application performs correlation evaluation on candidate multi-source covariates and historical global irrigation load, and then selects effective covariates according to the evaluation results; extracts a user feature vector based on the effective covariates; constructs a historical input sequence by combining the historical global irrigation load and the effective covariates, and constructs a future known covariate by combining the known future prior information in the prediction interval; constructs input data by combining the user feature vector, the historical input sequence and the future known covariate; inputs the input data into a load prediction model adopting a TimeXer framework, and outputs an irrigation load prediction result in the prediction interval.

Description

Technical Field

[0001] This application belongs to the field of power system load forecasting and intelligent analysis of agricultural irrigation energy consumption. Background Technology

[0002] Irrigation load mainly refers to the workload borne by irrigation-related equipment and systems during agricultural irrigation. Irrigation load is characterized by being "event-driven, intermittently operating, and significantly different among users," and is significantly affected by exogenous factors such as meteorological conditions (rainfall, radiation, temperature, humidity, and wind), crop growth stages, and irrigation regimes. Furthermore, different users exhibit significant differences in their irrigation time preferences, irrigation frequency, continuous operating time, and response patterns to exogenous conditions, leading to substantial differences in the shape of load curves among different users in the same region.

[0003] Existing short-cycle irrigation load prediction methods mainly include:

[0004] (1) Methods based on statistical regression or traditional machine learning: have limited ability to model exogenous factors and are difficult to capture nonlinearity and long-term dependence;

[0005] (2) Sequence models based on deep learning: Although they can fit complex time series, they often treat all users as homogeneous individuals and fail to effectively utilize user difference information;

[0006] (3) Methods that use only historical loads or only a few covariates: When exogenous drivers change (e.g., a sudden increase in rainfall, crop stage switching), forecasts are prone to lag or drift.

[0007] Therefore, there is a need for a short-cycle irrigation load forecasting scheme that can simultaneously utilize multi-source covariates (weather + crop stage, etc.) and user-specific characteristics, and can achieve conditional forecasting using a single model in multi-user scenarios. Summary of the Invention

[0008] This application aims to address the limitations of existing short-cycle irrigation load prediction methods in terms of their ability to model exogenous factors, their inability to effectively utilize user-specific information, and their susceptibility to prediction lag or drift. It presents a multi-source covariate short-cycle irrigation load prediction method based on the TimeXer framework. By integrating prior information from multiple covariates such as user differences, crop growth stages, and meteorological data, a conditional prediction model based on the TimeXer framework is constructed, enabling high-precision probabilistic interval prediction and risk assessment support for short-cycle irrigation loads for multiple users.

[0009] The first aspect of this application provides a method for short-cycle prediction of multi-source covariate irrigation load based on the TimeXer framework, including:

[0010] The correlation between candidate multi-source covariates and historical global irrigation load is evaluated, and effective covariates are then selected based on the evaluation results. The candidate multi-source covariates include meteorological time-series data and crop growth stage information.

[0011] Extract user feature vectors based on the effective covariates;

[0012] The historical global irrigation load and the effective covariate are used to construct a historical input sequence, and the known future prior information within the prediction interval is used to construct a future known covariate.

[0013] The user feature vector, historical input sequence, and known future covariates constitute the input data;

[0014] The input data is fed into a load forecasting model using the TimeXer framework, and the irrigation load forecast results for the forecast interval are output.

[0015] In one possible design, the correlation assessment of candidate multi-source covariates with historical irrigation load, and the subsequent selection of effective covariates based on the assessment results, includes:

[0016] Calculate the Pearson correlation coefficient between the candidate multi-source covariates and the historical global irrigation load;

[0017] The effective covariates were obtained by screening using the Pearson correlation coefficient.

[0018] In one possible design, calculating the Pearson correlation coefficient between the candidate multi-source covariates and the historical global irrigation load includes:

[0019] The Pearson correlation coefficient between the candidate multi-source covariates and the historical global irrigation load is calculated using the following formula:

[0020] ,

[0021] in, for and Pearson correlation coefficient, for At this moment One candidate covariate, for Real-time global irrigation load, The sample mean of the candidate covariates. This represents the mean of the irrigation load sample. This represents the total number of sample time points;

[0022] The process of using the Pearson correlation coefficient to screen for effective covariates includes:

[0023] Effective covariates are obtained by screening using the following formula:

[0024] ,

[0025] in, For an effective set of covariates, The total number of candidate covariates.

[0026] In one possible design, the extraction of user feature vectors based on the effective covariates includes:

[0027] The global statistical feature vector is characterized by total electricity consumption, event initiation frequency, average continuous running time, average irrigation stop interval duration, and time-segmented irrigation preferences.

[0028] Construct covariate response feature vectors using the response intensity of each user to each effective covariate;

[0029] The global statistical feature vector and the covariate response-related feature vector are concatenated to form the user feature vector.

[0030] In one possible design, the expression for the global statistical feature vector is:

[0031] ,

[0032] in, These are global statistical feature vectors;

[0033] Total electricity consumption within the statistical period: , For users exist The irrigation load at any time, This represents the set of time indices within a statistical period. The sampling interval;

[0034] Event trigger frequency: , For the number of events, For users exist The runtime indicator variable at any given time. Indicates that the program is running. Indicates that irrigation has been stopped;

[0035] The average duration of continuous operation: , This refers to the continuous runtime.

[0036] The average duration of irrigation interruptions: , This refers to the duration of the irrigation stoppage interval;

[0037] The daytime electricity consumption percentage is used to characterize users' irrigation preferences by time of day: , For daytime periods;

[0038] The expression for the covariate response feature vector is:

[0039] ,

[0040] in, For the covariate response feature vector, For users The irrigation load and the first Pearson correlation coefficients among the effective covariates Used to characterize users For the first The response strength of each effective covariate

[0041] ,

[0042] for At this moment One effective covariate, For the first The sample mean of each effective covariate within the statistical window. For users The mean of the irrigation load sample within the statistical window;

[0043] The expression for the user feature vector is:

[0044] ,

[0045] in, For users eigenvectors.

[0046] In one possible design, the expression for the historical input sequence is:

[0047] ,

[0048] in, For users At the predicted start time Historical input, For the length of the history window, Indicates user From time arrive The load sequence, Indicates from time arrive Effective covariates within the time period;

[0049] The expression for the known future covariates is:

[0050] ,

[0051] in, For a known sequence of covariates within a prediction interval, To predict the step size, To predict the start time Based on known covariate information, including: weather forecast values, stage-specific thermal codes determined according to agricultural time plans and growth stage rules.

[0052] In one possible design, the load forecasting model employing the TimeXer framework includes:

[0053] The endogenous variable processing module is used to extract features from historical irrigation load sequences to obtain endogenous feature representations that include local temporal features and global summary features.

[0054] The exogenous variable processing module is used to extract the overall impact features of the effective covariate sequence according to the variable dimensions to obtain exogenous feature representations.

[0055] The feature fusion module is used to use the global summary features in the endogenous feature representation as query items, and inject information from the exogenous feature representation into the endogenous feature representation through a cross-attention mechanism to obtain a comprehensive feature representation after fusing external factors.

[0056] The user condition embedding module is used to receive and embed user feature vectors, and introduce user difference conditions into the load prediction model.

[0057] The prediction output module is used to map the comprehensive feature representation and user-differential conditions into a prediction result of future irrigation load.

[0058] In one possible design, the objective function expression of the load forecasting model is:

[0059] ,

[0060] in, For the comprehensive training objectives, As a weighting factor, This represents the quantile loss function. and These represent the lower quantile and the upper quantile, respectively.

[0061] A second aspect of this application provides a short-cycle prediction device for multi-source covariate irrigation load based on the TimeXer framework. The device includes a processor and a memory. The memory stores at least one instruction, which is loaded and executed by the processor to implement the short-cycle prediction method for multi-source covariate irrigation load based on the TimeXer framework as described above.

[0062] A third aspect of this application provides a computer storage medium storing at least one instruction, which is loaded and executed by a processor to implement the multi-source covariate irrigation load short-cycle prediction method based on the TimeXer framework described above.

[0063] The beneficial effects of this application are:

[0064] (1) A multi-source covariate screening technique is proposed to construct a candidate set of meteorological covariates and crop growth stage covariates. The multivariate covariates are subjected to Pearson correlation analysis with historical irrigation load. Based on the principle that "the absolute value of the correlation is higher than the average value", an effective subset of covariates is selected to remove meteorological variables with weak correlation and retain exogenous information that has a significant driving effect on irrigation load.

[0065] (2) A user feature extraction technique is proposed, which extracts user feature vectors based on users' historical irrigation load data. The user features include at least: global statistical features reflecting energy consumption habits and irrigation event structure, and covariate response-related features reflecting users' sensitivity to and response patterns to selected covariates. A multi-source covariate system is constructed and time-aligned and feature-constructed. The correlation between the multi-source covariates and historical irrigation load is evaluated and a set of effective covariates is obtained. At the same time, user static feature vectors are extracted only based on the user's historical irrigation load sequence. The user static feature vectors are used to characterize the differences between users' irrigation event structure and energy consumption habits.

[0066] (3) A method for constructing an input dataset that integrates user static features and future prior information is proposed. In short-cycle prediction tasks, training samples are constructed based on a sliding window. User features are used as static covariates, historical irrigation load is used as the core time series input, and the selected multi-source covariates (meteorological + crop growth stage) are introduced into the historical segment and the future prior segment at the same time. A unified sample oriented towards the prediction starting point is constructed to form a conditional input dataset that can be shared and trained by multiple users.

[0067] (4) A conditional probability interval prediction model based on TimeXer is proposed. TimeXer is used as the main framework for time series prediction. By embedding user static features and integrating them with multi-source covariate inputs, personalized prediction for multiple users in a single model is achieved. At the same time, multi-quantile output or equivalent probability modeling is used to generate the prediction interval of future irrigation load, which is used to characterize prediction uncertainty and support risk assessment and scheduling decision. Attached Figure Description

[0068] Figure 1 Comparison chart of multi-step point predictions for short-cycle loads using the TimeXer model;

[0069] Figure 2 Comparison chart of multi-step interval prediction for short-cycle load using the TimeXer model;

[0070] Figure 3 This is a schematic diagram of the TimeXer model structure;

[0071] Figure 4 This is a flowchart of the method described in the embodiment;

[0072] Figure 5 Radar chart showing the prediction performance scores for each model. Detailed Implementation

[0073] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0074] In view of this, the embodiments of this application provide a multi-source covariate irrigation load short-cycle prediction method based on the TimeXer framework, in order to solve the above problems.

[0075] Specific Implementation Method 1: The multi-source covariate irrigation load short-cycle prediction method based on the TimeXer framework described in this implementation method includes:

[0076] The correlation between candidate multi-source covariates and historical global irrigation load is evaluated, and effective covariates are then selected based on the evaluation results. The candidate multi-source covariates include meteorological time-series data and crop growth stage information.

[0077] Extract user feature vectors based on the effective covariates;

[0078] The historical global irrigation load and the effective covariate are used to construct a historical input sequence, and the known future prior information within the prediction interval is used to construct a future known covariate.

[0079] The user feature vector, historical input sequence, and known future covariates constitute the input data;

[0080] The input data is fed into a load forecasting model using the TimeXer framework, and the irrigation load forecast results for the forecast interval are output.

[0081] In one implementation, the step of assessing the correlation between candidate multi-source covariates and historical irrigation load, and then selecting effective covariates based on the assessment results, includes:

[0082] Calculate the Pearson correlation coefficient between the candidate multi-source covariates and the historical global irrigation load;

[0083] The effective covariates were obtained by screening using the Pearson correlation coefficient.

[0084] In one implementation, calculating the Pearson correlation coefficient between the candidate multi-source covariates and the historical global irrigation load includes:

[0085] The Pearson correlation coefficient between the candidate multi-source covariates and the historical global irrigation load is calculated using the following formula:

[0086] ,

[0087] in, for and Pearson correlation coefficient, for At this moment One candidate covariate, for Real-time global irrigation load, The sample mean of the candidate covariates. This represents the mean of the irrigation load sample. This represents the total number of sample time points;

[0088] The process of using the Pearson correlation coefficient to screen for effective covariates includes:

[0089] Effective covariates are obtained by screening using the following formula:

[0090] ,

[0091] in, For an effective set of covariates, The total number of candidate covariates.

[0092] In one implementation, the step of extracting the user feature vector based on the effective covariates includes:

[0093] The global statistical feature vector is characterized by total electricity consumption, event initiation frequency, average continuous running time, average irrigation stop interval duration, and time-segmented irrigation preferences.

[0094] Construct covariate response feature vectors using the response intensity of each user to each effective covariate;

[0095] The global statistical feature vector and the covariate response-related feature vector are concatenated to form the user feature vector.

[0096] In one implementation, the expression for the global statistical feature sub-vector is:

[0097] ,

[0098] in, These are global statistical feature vectors;

[0099] Total electricity consumption within the statistical period: , For users exist The irrigation load at any time, This represents the set of time indices within a statistical period. The sampling interval;

[0100] Event trigger frequency: , For the number of events, For users exist The runtime indicator variable at any given time. Indicates that the program is running. Indicates that irrigation has been stopped;

[0101] The average duration of continuous operation: , This refers to the continuous runtime.

[0102] The average duration of irrigation interruptions: , This refers to the duration of the irrigation stoppage interval;

[0103] The daytime electricity consumption percentage is used to characterize users' irrigation preferences by time of day: , For daytime periods;

[0104] The expression for the covariate response feature vector is:

[0105] ,

[0106] in, For the covariate response feature vector, For users The irrigation load and the first Pearson correlation coefficients among the effective covariates Used to characterize users For the first The response strength of each effective covariate

[0107] ,

[0108] for At this moment One effective covariate, For the first The sample mean of each effective covariate within the statistical window. For users The mean of the irrigation load sample within the statistical window;

[0109] The expression for the user feature vector is:

[0110] ,

[0111] in, For users eigenvectors.

[0112] In one implementation, the expression for the historical input sequence is:

[0113] ,

[0114] in, For users At the predicted start time Historical input, For the length of the history window, Indicates user From time arrive The load sequence, Indicates from time arrive Effective covariates within the time period;

[0115] The expression for the known future covariates is:

[0116] ,

[0117] in, For a known sequence of covariates within a prediction interval, To predict the step size, To predict the start time Based on known covariate information, including: weather forecast values, stage-specific thermal codes determined according to agricultural time plans and growth stage rules.

[0118] In one implementation, the load forecasting model employing the TimeXer framework includes:

[0119] The endogenous variable processing module is used to extract features from historical irrigation load sequences to obtain endogenous feature representations that include local temporal features and global summary features.

[0120] The exogenous variable processing module is used to extract the overall impact features of the effective covariate sequence according to the variable dimensions to obtain exogenous feature representations.

[0121] The feature fusion module is used to use the global summary features in the endogenous feature representation as query items, and inject information from the exogenous feature representation into the endogenous feature representation through a cross-attention mechanism to obtain a comprehensive feature representation after fusing external factors.

[0122] The user condition embedding module is used to receive and embed user feature vectors, and introduce user difference conditions into the load prediction model.

[0123] The prediction output module is used to map the comprehensive feature representation and user-differential conditions into a prediction result of future irrigation load.

[0124] In one implementation, the objective function expression of the load forecasting model is:

[0125] ,

[0126] in, For the comprehensive training objectives, As a weighting factor, This represents the quantile loss function. and These represent the lower quantile and the upper quantile, respectively.

[0127] To further illustrate the implementation scheme of this application, Figure 4 A method for short-cycle prediction of multi-source covariate irrigation load based on the TimeXer framework is provided, comprising steps one through four. The numbering of these steps does not necessarily preclude their execution order. Each step is described in detail below:

[0128] Step 1: Obtain historical time-series data of irrigation loads from multiple users and their corresponding multi-source covariate data. Assess the correlation between the multi-source covariate data and the historical irrigation load data, and then select the effective covariate set. Details are as follows:

[0129] Multi-source covariate data mainly includes meteorological time-series data and crop growth stage information. Pearson correlation was calculated using the global irrigation load as a reference sequence, and a subset of effective covariates with strong correlation to irrigation load was selected, as shown in formulas (1) to (5):

[0130] Construct candidate multi-source covariate vectors:

[0131] (1),

[0132] in, for The candidate multi-source covariate vector at time step; for The time-based meteorological covariate subvector, which includes: rainfall, temperature, humidity, wind speed, radiation, and sunshine, etc. for The crop growth stage covariate subvectors at time points include: growth stage coding, number of days after sowing / transplanting, accumulated temperature, crop coefficient, evapotranspiration correlation, irrigation demand index, etc. This is a discrete-time index.

[0133] Construct a global irrigation load reference sequence:

[0134] (2),

[0135] in, for Global irrigation load reference sequence at time t, For users exist The irrigation load at any time, For the number of users.

[0136] Using formula (3), the crop growth stages are encoded as one-hot vectors that can be directly input into the model, thereby improving the interpretability of crop growth law variables:

[0137] (3),

[0138] in, This represents the unique heat vectors for crop growth stages. Represents the one-heat function. Encoding crop growth stages, This represents the number of growth stage categories.

[0139] Calculate the Pearson correlation coefficient between the candidate covariates and the global load using formula (4):

[0140] (4),

[0141] in, for and The Pearson correlation coefficient; For the first The candidate covariates are in The value at time; Represents the sample mean of the covariates; This represents the global load sample mean; This represents the total number of sample time points.

[0142] Formula (5) is used to set the mean threshold and obtain the filtered set of covariates and the final covariate vector:

[0143] (5),

[0144] ,

[0145] in, The set of covariates after filtering; The total number of candidate covariates; for The final covariate vector at time step; For the dimensions of the covariates after filtering; For the first Each covariate in The value at time.

[0146] Step 2: Based on users' historical irrigation load energy consumption habits and irrigation event structure characteristics, as well as users' historical sensitivity and response patterns to the effective covariates, extract user feature vectors as static input conditions for subsequent prediction models. Specifically:

[0147] Based on the user's historical irrigation load and its historical response relationship with the effective covariates obtained in step one, user difference features are extracted to form a user feature vector, which serves as the static input condition for the subsequent prediction model. The core process is described by formulas (6) to (10):

[0148] Define the overall representation of user profiles, decompose user characteristics into global statistical features and covariate response-related features, and then concatenate them to reflect not only differences in users' irrigation energy consumption habits and irrigation event structures, but also differences in users' sensitivity to exogenous conditions and response patterns.

[0149] (6),

[0150] in, For users eigenvectors; These are global statistical feature vectors; For the covariate response feature vector.

[0151] Using formula (7), the operating status of the user's historical irrigation load is determined, and the continuous load time series is converted into a binary sequence of operation / stoppage, forming an identifiable irrigation event structure:

[0152] (7),

[0153] in, For users exist The irrigation load at any given time; Thresholds for determining irrigation operation; For users exist The runtime indicator variable at any given time. Indicates that the program is running. This indicates that irrigation has been stopped.

[0154] Based on this, the continuous runtime is given. Duration of irrigation stoppage Calculation method:

[0155] (8),

[0156] in, and users respectively In the The start and end times of the next running event; The sampling interval; users respectively In the The duration of continuous operation and the duration of irrigation stoppage interval in each operation event.

[0157] Formula (8) provides quantifiable event-level indicators for the global statistical sub-vectors (frequency, running / interruption duration, etc.) obtained by summarizing Formula (9). Formula (9) is used to construct the user's global statistical feature sub-vectors, including total electricity consumption, start-up frequency, running and interruption duration statistics, and time-segment preferences, to characterize the user's irrigation scale and event intensity.

[0158] (9),

[0159] (10)

[0160] in, The total electricity consumption (or load integral) within the statistical period. This refers to the number of times the event is run (start frequency). This is the average duration of continuous operation; This represents the average duration of the irrigation stoppage interval; The percentage of electricity consumption during the day is used to characterize users' irrigation preferences at different times of the day. This represents the set of time indices within a statistical period. Indicates a gathering during the daytime period. This represents the number of events.

[0161] Formula (11) is used to construct the user covariate response feature subvector. By analyzing the historical correlation strength between user load and the covariates selected in step (1), it characterizes the differences in user sensitivity to exogenous information such as weather and crop growth stages, thereby providing user-differential input conditions for the prediction model.

[0162] (11),

[0163] (12)

[0164] in, For users Load and the first Pearson correlation coefficient among the covariates Used to characterize users For each effective covariate The response intensity, which in turn constitutes the user covariate response feature subvector. ; The multi-source covariate vector obtained in step (1) is... For the first Each covariate in The value at time, For covariate dimensions; For users The mean of the irrigation load sample within the statistical window For the first The sample mean of each effective covariate within the statistical window.

[0165] Step 3: Construct training samples based on a sliding window. These training samples should include at least historical irrigation load sequences, historical segments and known future segments of effective covariates, user feature vectors, and corresponding future irrigation load supervision labels, thus forming an input dataset that integrates user features and future prior information. Specifically:

[0166] This step proposes a method for constructing an input dataset that integrates user features and future prior information, using the user feature vector extracted in step two. As static conditional information (static covariates), the user's historical irrigation load and the multi-source covariates selected in step one are used to construct a historical input sequence; at the same time, the future prior information that can be obtained in advance within the prediction interval (such as weather forecasts, crop growth stage unique heat codes, etc.) are used to construct "future known covariates" input, and the future irrigation load sequence is used as a supervision label, thereby forming training samples that can be used for short-cycle multi-step prediction. Its core process is described by formulas (14) to (17):

[0167] Constructing the historical input sequence:

[0168] (14)

[0169] in, For users At the predicted start time Historical input; The length of the history window; Indicates from time arrive The load sequence; Indicates from time arrive The multi-source covariate sequence within the time period (obtained from step one).

[0170] Construct known covariates for the future:

[0171] (15)

[0172] in, To predict the future sequence of known covariates within a given interval; To predict the step size; This refers to covariate information that is available at the moment the forecast is initiated (such as weather forecast values, stage-specific thermal codes determined according to agricultural time plans / growth stage rules, etc.).

[0173] Constructing supervision labels:

[0174] (16)

[0175] in, To supervise learning labels, i.e., the future The actual irrigation load sequence of the step is also the prediction target of the model.

[0176] Based on formulas (14), (15), and (16), the final structure of a single sample (user features + historical input + future known input + supervision label) is given:

[0177] (17)

[0178] in, This represents a training sample; The user feature vectors (static covariates) obtained in step two are used to provide user difference conditions to the model.

[0179] After completing step three, "Constructing an input dataset that integrates user features and future prior information," to ensure the method can be directly implemented and interfaced with the prediction model in step four, this embodiment organizes each training sample into a unified format of "endogenous sequence input, exogenous sequence input, and user static feature input." The historical irrigation load sequence serves as the endogenous input, and to enhance the expression of local variation patterns and long-term dependencies, it can be divided into several time segments of fixed length and represented accordingly. The multi-source covariates obtained in step one serve as the exogenous input, constructed by concatenating historical covariates with known future covariates (e.g., weather forecast values, stage-specific thermal codes determined according to agricultural time plans / growth stage rules) within the prediction interval, and organized according to "variable dimensions" to characterize the differences in the impact of different external factors on irrigation load. The user static feature vector in step two... As static conditional information, it, along with the aforementioned endogenous and exogenous inputs, constitutes the model input, used to provide user-differentiated conditions to the unified prediction model. Based on the above sample organization method, step four realizes the interaction and fusion of endogenous and exogenous information within the model, and outputs the future... The output can be expanded to include upper and lower bound prediction sequences to meet the needs of short-cycle interval prediction.

[0180] Step 4: Input the input dataset into the short-cycle load prediction model for training and inference. The prediction model uses the TimeXer framework to fuse and model historical irrigation loads and multi-source covariates, outputting prediction results for future multi-step irrigation load points, and extending quantile learning at the output end to simultaneously provide the upper and lower bounds of the prediction interval.

[0181] This embodiment treats irrigation load as an endogenous variable to be predicted, and the selected meteorological variables and crop growth stages as exogenous covariates. The prediction task can be expressed as: given a user Historical irrigation load sequences and multi-source covariate sequences are used to output future... The irrigation load forecast for this step is consistent with TimeXer's "exogenous variable assists endogenous variable prediction" setting. Additionally, this embodiment introduces user feature vectors. This is used to reflect user differences. Therefore, the final prediction form of this embodiment can be expanded to:

[0182] (18)

[0183] in, For the predicted results, Historical inputs (including historical loads and historical covariates); For known future covariates that can be obtained in advance within the prediction interval (such as weather forecasts, crop growth stage plans, etc.); This is a user feature vector, used to express the differences in energy usage habits and external response conditions among different users. This represents the prediction function.

[0184] Specifically, in order to enhance the expression of the local regularity and long-term dependence of irrigation load in the "event-driven, intermittent operation" mode, this embodiment adopts a fragmented representation of the historical irrigation load sequence: the continuous historical sequence is divided into several time segments of fixed length, and the features of each segment are extracted. At the same time, a global summary feature that can summarize the entire history is constructed. Based on the above segment features and global summary features, they are jointly updated by the internal interactive update module to obtain an endogenous feature representation that integrates local and global information, corresponding to formula (19).

[0185] (19)

[0186] in, Show the first The layer's "load segment feature representation" is used to describe the characteristics of historical irrigation loads at different time segments; Indicates the first The layer "global summary feature representation" is used to summarize the overall information of the entire historical irrigation load; This means that the fragment features and global features are concatenated and then used as a unified feature sequence as input. This is a self-attention module used to establish relationships between fragments and between fragments and the global context, and to output update values. "+" indicates normalization, used to improve the stability of multi-layer stacked training; "+" indicates residual stacking, which adds the original features to the update amount to retain the original information and enhance training stability.

[0187] Multi-source covariates are represented at the variable level: their overall influence features are extracted according to the covariate dimension, thereby reducing the dependence on strict time-by-time alignment; in the fusion stage, the global summary features of the endogenous sequence are used as the fusion bridge to summarize the information of external covariates and fuse them with the endogenous feature representation to form the prediction basis after fusing external factors, corresponding to formula (20):

[0188] (20)

[0189] In the formula: Indicates the first The layer is used for global feature representation in external information fusion. Indicates the sample index; A set of features representing external covariates, used to provide information about external influences; For cross-attention module ( Create a query. (Used as a key / value pair) to inject aggregated external covariate information into the global feature representation.

[0190] Endogenous output embedding after obtaining fused exogenous information Then, the final endogenous representation needs to be mapped to the future. The invention provides the irrigation load prediction results and the corresponding training objective function. TimeXer's original approach is to linearly project the endogenous output embedding to obtain the point prediction sequence, with exogenous variables serving only as auxiliary information inputs and not as prediction outputs. Based on this, the invention adds a quantile output head at the projection end to further obtain upper and lower bound prediction sequences, and uses quantile loss to constrain the reliability of interval prediction, thereby achieving integrated training of point prediction and interval prediction.

[0191] Formula (21) is used to characterize the output form of point prediction and the training objective of squared loss.

[0192] (twenty one),

[0193] in, This represents a linear mapping to the final layer's endogenous representation, outputting the future... Step-by-step predicted value; Used to measure the deviation between the predicted sequence and the true sequence. For the future Step-point prediction sequence, For the first Step into the actual load, For the first Predicting load step by step; Embed the endogenous output of the final layer of TimeXer.

[0194] Formula (22) is used to characterize the pinball loss of interval prediction, which enables the model to learn the conditional distribution of a specified quantile by asymmetric penalty for different quantiles.

[0195] (twenty two),

[0196] in, quantiles Next Step-by-step predicted value For the corresponding quantile loss, and These represent the lower quantile and the upper quantile, respectively (e.g., 0.1 and 0.9).

[0197] Formula (23) is used to give the comprehensive training objective, which weights and fuses the point prediction loss and the upper and lower quantile losses to balance prediction accuracy and interval reliability.

[0198] (twenty three),

[0199] in, For the comprehensive training objectives, This is a weighting factor used to balance the accuracy of point prediction and the reliability of interval prediction. Through the above output and training objective design, this embodiment further obtains the probability interval of future loads based on TimeXer's "endogenous-exogenous fusion prediction," reserving a safety margin for irrigation scheduling.

[0200] After the above steps, short-cycle point prediction and interval prediction of irrigation load can be achieved by integrating user differences and multi-source covariates.

[0201] Case Analysis

[0202] 1. Experimental environment and prediction model configuration

[0203] The computer used for the case study analysis was configured with an AMD Ryzen 9 5900HX processor with Radeon Graphics, 16.0GB of RAM, and an NVIDIA GeForce RTX 3060 Laptop GPU with 4GB of VRAM. The PyTorch 2.5.1+cu121 deep learning framework was used, with CUDA version 12.1, and the GPU was utilized for CUDA parallel computing to accelerate training.

[0204] 2. Baseline Model Comparison Experiment

[0205] To verify the effectiveness of the short-cycle load forecasting method described in this invention in terms of accuracy and robustness, this embodiment sets up a comparative experiment, selecting three representative baseline models: ARIMA, SVR, and LSTM. ARIMA represents a traditional statistical time series model, suitable for linear modeling of stationary or weakly non-stationary sequences; SVR represents a traditional machine learning regression model, adept at learning nonlinear mappings under small sample conditions; and LSTM represents a deep learning sequence model, capable of characterizing the long-term and short-term dependencies of time series. Through the comparison of these three baseline models, the advantages of the method described in this invention can be verified from different technical routes—"statistical modeling—machine learning—deep learning"—in scenarios such as complex nonlinearity, mutation / event-driven scenarios, and multi-step prediction error accumulation.

[0206] To ensure fairness in the comparison, the baseline model adopts a consistent data partitioning method and evaluation metric settings: the same training / test set partition is used for the same dataset, and a sliding window is uniformly used to construct samples, with the historical window length set to L and the prediction step size set to H. The baseline model only uses historical irrigation load sequences for training and prediction, without introducing exogenous covariates such as meteorological conditions and crop growth stages, or user characteristic variables; that is, the baseline model is strictly limited to the "univariate historical sequence prediction" setting in terms of information input. The evaluation metrics used are MAE, MSE, RMSE, and sMAPE. The method in this embodiment further introduces user-specific features and multi-source covariates under the same training / testing conditions to achieve conditional modeling of the load formation mechanism, thereby improving prediction accuracy and stability.

[0207] Baseline Model 1: ARIMA (Autoregressive Integral Moving Average)

[0208] Model Principle: ARIMA is a classic time series model, composed of Autoregression (AR), Difference (I), and Moving Average (MA). These are used to characterize the linear autocorrelation of the series, eliminate trend to make the series approximately stationary, and the short-term correlation of the residuals, respectively. The model form is usually denoted as... ,in It is the difference order. and These are the orders of AR and MA, respectively.

[0209] Baseline Model 2: SVR (Support Vector Regression)

[0210] Model principle: SVR is an extension of Support Vector Machines for regression tasks, which introduces... The insensitive loss and margin maximization approach can fit the input-output relationship while ensuring generalization ability; and the input can be mapped to a high-dimensional feature space through kernel functions (such as RBF kernel) to express nonlinear relationships.

[0211] Baseline Model 3: LSTM (Long Short-Term Memory Network)

[0212] Model principle: LSTM is an improved recurrent neural network that controls the retention and forgetting of information in the time dimension through the gating mechanism of "input gate, forget gate, output gate", thereby alleviating the gradient vanishing problem of traditional RNNs and enabling it to learn dependencies and nonlinear dynamic patterns over a long time span.

[0213] Because the data contains a large number of 0 or near-0 values, MAPE errors can be abnormally amplified due to the small denominator, resulting in distorted evaluation results. Therefore, this invention uses sMAPE, which normalizes the data using a combined scale of the true and predicted values. This avoids error explosion caused by zero values ​​and reflects the prediction accuracy more stably.

[0214] Table 1. Comparison of Prediction Results

[0215]

[0216] The results in the table show that, under the same training / test split and prediction step size settings, TimeXer achieves the best performance in all four metrics (MAE=1.0316, RMSE=1.7540, sMAPE=1.5274), outperforming baseline models such as ARIMA, SVR, and LSTM. ARIMA, limited by its linear assumptions, is insufficient in characterizing nonlinear fluctuations and abrupt changes in the load sequence. SVR is sensitive to sliding window feature construction and hyperparameters, and multi-step prediction is prone to error propagation. While LSTM can learn nonlinear dependencies, its stability and generalization ability are affected when the sample size is limited or the sequence is highly intermittent (with a large number of 0 / near-zero values), resulting in higher error metrics. Furthermore, given the large number of 0 or near-zero values ​​in the sequence, traditional MAPE is prone to evaluation distortion due to an excessively small denominator. This embodiment uses sMAPE to improve the stability and comparability of the error measurement.

[0217] Figure 5 Using each metric as an axis (MAE, MSE, RMSE, sMAPE), each curve represents a model's overall performance across these metrics. Currently, z-score standardization is used, and the error metric has been inversely processed; therefore, a larger value indicates a better model performance on that metric. The further out the curve and the higher the score, the smaller the error and the better the model's performance; a significant contraction on a particular axis indicates a relatively weaker model on that metric.

[0218] Specific Implementation Method Two: The multi-source covariate irrigation load short-cycle prediction device based on the TimeXer framework described in this implementation method includes a processor and a memory. The memory stores at least one instruction, which is loaded and executed by the processor to implement the multi-source covariate irrigation load short-cycle prediction method based on the TimeXer framework as described in Specific Implementation Method One.

[0219] Specific Implementation Method 3: A computer storage medium as described in this embodiment stores at least one instruction, which is loaded and executed by a processor to implement the multi-source covariate irrigation load short-cycle prediction method based on the TimeXer framework as described in Specific Implementation Method 1.

[0220] While specific embodiments of this application have been described herein with reference to them, it should be understood that these embodiments are merely examples of the principles and applications of this application. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of this application as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for short-cycle prediction of multi-source covariate irrigation load, characterized in that, include: The correlation between candidate multi-source covariates and historical global irrigation load is evaluated, and effective covariates are then selected based on the evaluation results. The candidate multi-source covariates include meteorological time-series data and crop growth stage information. Extract user feature vectors based on the effective covariates; The historical global irrigation load and the effective covariate are used to construct a historical input sequence, and the known future prior information within the prediction interval is used to construct a future known covariate. The user feature vector, historical input sequence, and known future covariates constitute the input data; The input data is fed into a load forecasting model using the TimeXer framework, and the irrigation load forecast results for the forecast interval are output.

2. The method for short-cycle prediction of multi-source covariate irrigation load according to claim 1, characterized in that, The process of evaluating the correlation between candidate multi-source covariates and historical irrigation load, and then selecting effective covariates based on the evaluation results, includes: Calculate the Pearson correlation coefficient between the candidate multi-source covariates and the historical global irrigation load; The effective covariates were obtained by screening using the Pearson correlation coefficient.

3. The method for short-cycle prediction of multi-source covariate irrigation load according to claim 2, characterized in that, The calculation of the Pearson correlation coefficient between the candidate multi-source covariates and the historical global irrigation load includes: The Pearson correlation coefficient between the candidate multi-source covariates and the historical global irrigation load is calculated using the following formula: ， in, for and Pearson correlation coefficient, for At this moment One candidate covariate, for Real-time global irrigation load, The sample mean of the candidate covariates. This represents the mean of the irrigation load sample. This represents the total number of sample time points; The process of using the Pearson correlation coefficient to screen for effective covariates includes: Effective covariates are obtained by screening using the following formula: ， in, For an effective set of covariates, The total number of candidate covariates.

4. The method for short-cycle prediction of multi-source covariate irrigation load according to claim 1, characterized in that, The extraction of user feature vectors based on the effective covariates includes: The global statistical feature vector is characterized by total electricity consumption, event initiation frequency, average continuous running time, average irrigation stop interval duration, and time-segmented irrigation preferences. Construct covariate response feature vectors using the response intensity of each user to each effective covariate; The global statistical feature vector and the covariate response-related feature vector are concatenated to form the user feature vector.

5. The method for short-cycle prediction of multi-source covariate irrigation load according to claim 4, characterized in that, The expression for the global statistical feature sub-vector is: ， in, These are global statistical feature vectors; Total electricity consumption within the statistical period: , For users exist The irrigation load at any time, This represents the set of time indices within a statistical period. The sampling interval; Event trigger frequency: , For the number of events, For users exist The runtime indicator variable at any given time. Indicates that the program is running. Indicates that irrigation has been stopped; The average duration of continuous operation: , This refers to the continuous runtime. The average duration of irrigation interruptions: , This refers to the duration of the irrigation stoppage interval; The daytime electricity consumption percentage is used to characterize users' irrigation preferences by time of day: , For daytime periods; The expression for the covariate response feature vector is: ， in, For the covariate response feature vector, For users The irrigation load and the first Pearson correlation coefficients among the effective covariates Used to characterize users For the first The response strength of each effective covariate ， for At this moment One effective covariate, For the first The sample mean of each effective covariate within the statistical window. For users The mean of the irrigation load sample within the statistical window; The expression for the user feature vector is: ， in, For users eigenvectors.

6. The method for short-cycle prediction of multi-source covariate irrigation load according to claim 1, characterized in that, The expression for the historical input sequence is: ， in, For users At the predicted start time Historical input, For the length of the history window, Indicates user From time arrive The load sequence, Indicates from time arrive Effective covariates within the time period; The expression for the known future covariates is: ， in, For a known sequence of covariates within a prediction interval, To predict the step size, To predict the start time Based on known covariate information, including: weather forecast values, stage-specific thermal codes determined according to agricultural time plans and growth stage rules.

7. The method for short-cycle prediction of multi-source covariate irrigation load according to claim 1, characterized in that, The load forecasting model using the TimeXer framework includes: The endogenous variable processing module is used to extract features from historical irrigation load sequences to obtain endogenous feature representations that include local temporal features and global summary features. The exogenous variable processing module is used to extract the overall impact features of the effective covariate sequence according to the variable dimensions to obtain exogenous feature representations. The feature fusion module is used to use the global summary features in the endogenous feature representation as query items, and inject information from the exogenous feature representation into the endogenous feature representation through a cross-attention mechanism to obtain a comprehensive feature representation after fusing external factors. The user condition embedding module is used to receive and embed user feature vectors, and introduce user difference conditions into the load prediction model. The prediction output module is used to map the comprehensive feature representation and user-differential conditions into a prediction result of future irrigation load.

8. The method for short-cycle prediction of multi-source covariate irrigation load according to claim 1, characterized in that, The objective function expression of the load forecasting model is: ， in, For the comprehensive training objectives, As a weighting factor, This represents the quantile loss function. and These represent the lower quantile and the upper quantile, respectively.

9. A multi-source covariate irrigation load short-cycle prediction device based on the TimeXer framework, characterized in that, The multi-source covariate irrigation load short-cycle prediction device based on the TimeXer framework includes a processor and a memory. The memory stores at least one instruction, which is loaded and executed by the processor to implement the multi-source covariate irrigation load short-cycle prediction method as described in any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The computer storage medium stores at least one instruction, which is loaded and executed by a processor to implement the multi-source covariate irrigation load short-cycle prediction method as described in any one of claims 1 to 8.

Images