Power load prediction method and device based on order driving, medium and product

Abstract

The invention provides a power load prediction method and device based on order driving, a medium and a product, and the method comprises the following steps: obtaining historical power load data and production business data corresponding to a target time period, the production business data comprising an order material list and a workshop production scheduling plan of a manufacturing execution system; according to the process operation time window and the equipment energy consumption reference, constructing a pre-scheduling production feature sequence aligned with the historical load time; generating a time sequence input tensor through multi-dimensional splicing, and inputting the time sequence input tensor into the recurrent neural network model to obtain an initial predicted value; and monitoring the time deviation between the actual production progress and the production scheduling plan in real time, correcting the feature sequence based on the actual progress and inputting the model again when the deviation exceeds a threshold value, and outputting a corrected predicted value. By implementing the technical scheme provided by the invention, dynamic coupling of load prediction and production rhythm is realized, and the prediction precision is remarkably improved.

Description

Technical Field

[0001] This application relates to the field of industrial energy management technology, and in particular to an order-driven power load forecasting method, device, medium, and product. Background Technology

[0002] How to accurately predict the power load in industrial production processes, so as to precisely guide production scheduling and power purchase decisions, has become a core requirement for ensuring the balance of power grid supply and demand and reducing enterprise operating costs.

[0003] In existing technologies, industrial power load forecasting is typically achieved through pure time-series analysis based on historical electricity consumption data. For example, average load values ​​from the same historical period or statistical models trained on historical data are used to extrapolate future electricity consumption trends. However, existing technologies primarily rely on the statistical regularities of historical load data for extrapolation, failing to consider real-time order details and production schedules within the enterprise. This results in a disconnect between the load forecasting model and the actual energy consumption rhythm of production. When faced with frequently changing production plans in discrete manufacturing enterprises, such models cannot adjust their forecasts in real time by sensing dynamic changes on the business side. Consequently, the forecast curve struggles to accurately follow actual load fluctuations driven by specific orders. This leads to insufficient forecast accuracy when enterprises engage in refined energy dispatching or participate in electricity spot market transactions, potentially resulting in high deviation assessments and economic losses. Summary of the Invention

[0004] In view of this, this application provides an order-driven power load forecasting method, device, medium, and product to solve the above problems.

[0005] Firstly, an order-driven power load forecasting method is provided, which includes:

[0006] Acquire historical power load data prior to the target forecast period, as well as production business data corresponding to the target forecast period. The production business data includes the order bill of materials and the workshop scheduling plan from the manufacturing execution system.

[0007] Based on the operation time window of each process in the workshop production schedule and the equipment energy consumption benchmark corresponding to the order material list, a pre-scheduled production feature sequence aligned with historical power load data in terms of time resolution is constructed.

[0008] Historical power load data and pre-scheduled production characteristic sequences are spliced ​​together using multi-dimensional features to generate a time-series input tensor.

[0009] The time-series input tensor is input into a pre-trained recurrent neural network prediction model to obtain the initial power load prediction value for the target prediction period.

[0010] Real-time monitoring of the actual production progress of the manufacturing execution system, and calculation of the time deviation between the actual production progress and the workshop production schedule;

[0011] When the time deviation exceeds the preset adjustment threshold, the pre-scheduled production feature sequence is corrected based on the actual production progress, and the corrected pre-scheduled production feature sequence is re-inputted into the recurrent neural network prediction model to output the corrected power load prediction value.

[0012] The above technical solution, by combining the order bill of materials and workshop production schedule with historical power load data, can fully consider the planned and real-time nature of production operations in load forecasting, enabling the forecast results to dynamically reflect the actual impact of production activities on energy consumption. When the actual production progress deviates from the plan, the load can be re-forecasted by correcting the pre-scheduled production characteristic sequence, which can effectively eliminate the forecasting error caused by scheduling deviation and improve the real-time accuracy and adaptability of power load forecasting.

[0013] Optionally, based on the operation time windows of each process in the workshop production schedule and the equipment energy consumption benchmarks corresponding to the order bill of materials, a pre-scheduled production feature sequence aligned with historical power load data in terms of time resolution is constructed, specifically including:

[0014] Parse the material code information contained in the order bill of materials, and query the production equipment identifier associated with the material code information and the rated power parameter corresponding to the production equipment identifier in the preset equipment energy consumption database;

[0015] Analyze the workshop production schedule and extract the start and stop times and operating duration of the production equipment corresponding to each process.

[0016] Theoretical energy consumption curves are generated based on rated power parameters, start-up and stop times, and operating duration.

[0017] The theoretical energy consumption curve is resampled and interpolated according to the time sampling frequency of historical power load data to generate a pre-scheduled production characteristic sequence.

[0018] The above technical solution generates theoretical energy consumption curves by analyzing the order bill of materials and production schedule, combined with the equipment energy consumption database, and resamples and interpolates at the same sampling frequency as historical load data. This allows for the construction of a refined production energy consumption feature sequence in the time dimension, enabling the prediction model to accurately capture the start-up and shutdown status and power changes of equipment in each process, thereby achieving a high-precision correspondence between power load and production behavior.

[0019] Optionally, historical power load data and pre-scheduled production characteristic sequences are multi-dimensionally concatenated to generate a time-series input tensor, specifically including:

[0020] Outlier detection and removal are performed on historical power load data, and the resulting data gaps are filled using Lagrange interpolation.

[0021] The max-min normalization algorithm is used to map the filled historical power load data and the pre-scheduled production characteristic sequence to a preset numerical range;

[0022] Extract calendar feature data corresponding to the target prediction period. The calendar feature data should include at least weekday identifiers, holiday identifiers, and seasonal index identifiers.

[0023] In terms of feature dimension, normalized historical power load data, normalized pre-scheduled production feature sequence, and calendar feature data are concatenated in parallel to construct a three-dimensional time-series input tensor.

[0024] The above technical solutions, by removing outliers and interpolating historical load data, and by normalizing and multi-dimensionally stitching all feature data, can effectively eliminate interference caused by differences in units, ensuring that data from different sources are integrated in the same numerical space. Simultaneously, the introduction of calendar feature data enables the prediction model to identify the impact of external periodic factors such as weekdays, holidays, and seasons on power load, thereby improving the prediction model's ability to represent multi-source data.

[0025] Optionally, the time-series input tensor is input into a pre-trained recurrent neural network prediction model to obtain the initial power load prediction value for the target prediction period, specifically including:

[0026] The temporal input tensor is fed into the input layer of the recurrent neural network prediction model, and the input layer converts the temporal input tensor into a hidden layer feature vector.

[0027] The hidden layer feature vectors are processed through the long short-term memory network unit in the recurrent neural network prediction model to output the hidden state vector;

[0028] The hidden state vector is input into the fully connected layer of the recurrent neural network prediction model and linearly transformed to obtain the initial power load prediction value corresponding to the time step of the target prediction period.

[0029] The above technical solution, by inputting a time-series input tensor into a recurrent neural network and utilizing long short-term memory units to process the input features, can capture the temporal correlation and long-term dependency between power load and production characteristics. The prediction results output after linear transformation by a fully connected layer can reflect the complex nonlinear dynamic changes in the time series, achieving high-precision modeling and prediction of power load for the target prediction period.

[0030] Optionally, before inputting the temporal input tensor into the pre-trained recurrent neural network prediction model, the method further includes a training step for constructing the recurrent neural network prediction model, the training step specifically including:

[0031] Acquire historical measured load data, historical actual production scheduling records, and historical calendar feature data within a preset historical time period;

[0032] Characterize the historical actual production scheduling records to construct a historical production feature sequence that is aligned with the historical measured load data in the time dimension;

[0033] Historical measured load data, historical production feature sequences, and historical calendar feature data are concatenated along the feature dimension to generate a historical training sample set.

[0034] The historical training sample set is fed into the recurrent neural network prediction model to be trained. The loss function value between the model output and the historical measured load data is calculated. The weight parameters of the recurrent neural network prediction model to be trained are updated using an optimization algorithm until the loss function value meets the preset convergence condition, thus obtaining the pre-trained recurrent neural network prediction model.

[0035] The above technical solution constructs a training sample set containing historical load, production scheduling, and calendar features to supervise the training of a recurrent neural network, enabling the model to learn the mapping relationship between production plans, energy consumption patterns, and load changes. By continuously adjusting the weight parameters using a loss function and optimization algorithm, a predictive model with strong generalization ability under different production scenarios can be obtained, providing a stable and reliable foundation for subsequent predictions.

[0036] Optionally, the historical actual production scheduling records are characterized to construct a historical production feature sequence aligned with the historical measured load data in the time dimension, specifically including:

[0037] Remove abnormal records from the historical actual production scheduling records that are not aligned with the timestamps of the historical measured load data;

[0038] Analyze the production equipment operation data contained in the historical actual production scheduling records, and generate historical energy consumption curves based on the production equipment operation data and preset rated power parameters;

[0039] The historical energy consumption curve is resampled and interpolated according to the time sampling frequency of the historical measured load data to generate a historical production characteristic sequence.

[0040] The maximum-minimum normalization algorithm is used to map the historical production characteristic sequence to a preset numerical range that is consistent with the historical measured load data.

[0041] The above technical solutions, by cleaning, resampling, and normalizing historical production scheduling records, eliminate the noise caused by inconsistent timestamps and missing data, ensuring that the generated historical production feature sequence is completely aligned with historical load data in the time dimension. Combining the energy consumption curve generated from the equipment's rated power with a unified numerical range guarantees the integrity and consistency of the training data, thereby improving the stability of model training and the accuracy of prediction results.

[0042] Optionally, when the time deviation exceeds a preset adjustment threshold, the pre-scheduled production characteristic sequence is corrected based on the actual production progress, specifically including:

[0043] The lag or advance time of the actual production progress relative to the workshop production schedule is determined based on the positive or negative attribute of the time deviation value.

[0044] Locate the feature data segment corresponding to the unexecuted process after the current moment in the pre-scheduled production feature sequence;

[0045] Based on the lag or advance time, the feature data segment is shifted along the time axis to generate the shifted intermediate feature sequence.

[0046] The missing time windows caused by the translation of the intermediate feature sequence are filled with zero values ​​or filled with the mean value based on the feature values ​​of adjacent time steps to obtain the corrected pre-schedule production feature sequence.

[0047] The above technical solution determines whether production progress is ahead or behind based on time deviations and shifts and corrects the pre-scheduled feature sequence along the time axis, enabling the feature sequence to match the actual production status in real time. The missing time windows after shifting are appropriately filled to avoid model input distortion caused by feature breakage, thereby achieving rapid adaptive correction of the prediction model when production rhythm changes, improving the real-time response capability and dynamic accuracy of load forecasting.

[0048] In a second aspect, an electronic device is provided, including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any of the above.

[0049] Thirdly, a computer-readable storage medium is provided that stores instructions which, when executed, perform the method as described in any of the preceding descriptions.

[0050] Fourthly, a computer program product containing instructions is provided, which, when run on a server, causes the server to perform the method described in the first aspect and any possible implementation thereof.

[0051] Understandably, the electronic device provided in the second aspect, the computer-readable storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the method provided in this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0052] In summary, implementing one or more technical solutions provided in this application has at least the following technical effects or advantages:

[0053] By incorporating order bills of materials and workshop production schedules into the power load forecasting process, this method deeply integrates enterprise production data with energy consumption characteristics, enabling a shift from a historical statistics-driven to a business logic-driven forecasting model. This approach allows for the inclusion of production planning information in the forecasting phase, making the forecasts more forward-looking and reflecting future energy consumption trends in production activities. Furthermore, by integrating the real-time production progress monitoring mechanism of the Manufacturing Execution System (MES), a dynamic self-correction mechanism for the forecasting model is established. This enables the model to automatically adjust forecast results based on changes in production pace, improving the robustness and real-time responsiveness of the forecasts.

[0054] Furthermore, by constructing a production feature sequence with a temporal resolution consistent with historical load data, the energy consumption variation patterns of processes can be characterized at a fine-grained level, thereby improving the model's accuracy in identifying short-cycle load fluctuations. Utilizing multi-dimensional feature fusion and a recurrent neural network structure, complex temporal dependencies and interaction features between multi-source data can be effectively extracted, making the prediction results more stable, continuous, and closely aligned with actual operating conditions. Overall, this application can provide industrial enterprises with refined and dynamic energy consumption prediction support, contributing to the optimization of production plans, the balancing of power resource scheduling, and the intelligent management of energy costs. Attached Figure Description

[0055] Figure 1 This is an exemplary system architecture diagram of an order-driven power load forecasting method disclosed in this application;

[0056] Figure 2 This is a flowchart illustrating an order-driven power load forecasting method disclosed in this application;

[0057] Figure 3 This is a schematic diagram of the structure of an electronic device disclosed in this application.

[0058] Explanation of reference numerals in the attached figures: 100, System architecture; 101, First terminal device; 102, Second terminal device; 103, Third terminal device; 104, Network; 105, Server; 301, Processor; 302, Communication bus; 303, User interface; 304, Network interface; 305, Memory. Detailed Implementation

[0059] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0060] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0061] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0062] Figure 1 An exemplary system architecture diagram is shown, illustrating an embodiment of an order-driven power load forecasting method applicable to this application.

[0063] like Figure 1 As shown, the system architecture 100 may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used as a medium to provide communication links between the terminal devices 101, 102, 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0064] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as model training applications, video recognition applications, web browser applications, social platform software, etc.

[0065] Terminal devices 101, 102, and 103 can be either hardware or software. When terminal devices 101, 102, and 103 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, e-book readers, MP3 (Moving Picture Experts Group Audio Layer III) players, MP4 (Moving Picture Experts Group Audio Layer IV) players, laptops, and desktop computers, etc. When terminal devices 101, 102, and 103 are software, they can be installed in the aforementioned electronic devices. They can be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules used to provide distributed services) or as a single software program or software module. No specific limitations are imposed here.

[0066] When terminals 101, 102, and 103 are hardware devices, video capture devices can also be installed on them. These video capture devices can be various devices capable of capturing video, such as cameras, sensors, etc. Users can use the video capture devices on terminals 101, 102, and 103 to capture video.

[0067] Server 105 can be a server that provides various services, such as a backend server for processing data displayed on terminal devices 101, 102, and 103. The backend server can analyze and process the received data and can feed back the processing results (such as recognition results) to the terminal devices.

[0068] It should be noted that a server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules (e.g., multiple software programs or software modules used to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0069] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included. In particular, if the target data does not need to be obtained remotely, the above system architecture may exclude the network and include only terminal devices or servers.

[0070] Figure 2This is a flowchart illustrating an order-driven power load forecasting method according to an embodiment of this application. This method can be implemented using a computer program or a microcontroller. The computer program can be integrated into an application or run as a standalone utility application. The specific steps of an order-driven power load forecasting method are described in detail below.

[0071] S201: Obtain historical power load data prior to the target forecast period, as well as production business data corresponding to the target forecast period. The production business data includes the order bill of materials and the workshop scheduling plan from the manufacturing execution system.

[0072] In this embodiment, production business data refers to business-level information that reflects the actual operational actions and energy consumption in the industrial production process. It is used to characterize the internal production activity arrangements and resource demand status of an enterprise over a future period. For example, the production business data mainly consists of static attribute data that determines "what to produce" and dynamic scheduling data that determines "when to produce." For example, it may include, but is not limited to, the Bill of Materials (BOM) exported from the Enterprise Resource Planning (ERP) system and the shop floor production schedule obtained from the Manufacturing Execution System (MES).

[0073] Specifically, the system needs to define the time range for power load forecasting, i.e., the target forecast period (e.g., the next 24 hours or the next week). For historical periods preceding this target forecast period, historical power load data is retrieved via the API (Application Programming Interface) of smart meters, Supervisory Control and Data Acquisition (SCADA) systems, or the enterprise energy management database. This data is typically represented as time-series power values ​​recorded at a fixed sampling frequency (e.g., every 15 minutes or every hour), reflecting the enterprise's basic energy consumption habits, cyclical fluctuations, and baseline levels of non-productive loads in past production cycles. Simultaneously, to capture the dynamic incremental impact of future production actions on power load, business data covering the target forecast period needs to be obtained across systems. This includes accessing the Enterprise Resource Planning (ERP) system to extract the Bill of Materials (BOM) based on active orders within this period, thereby determining the specifications and required raw materials of the products to be produced; and accessing the Manufacturing Execution System (MES) to obtain workshop scheduling plans accurate to the process level. These plans detail the specific equipment identifiers, planned start and end times, and expected output for each production task. By aggregating the aforementioned historical energy consumption records with future production and business intentions, a complete data foundation is provided for the subsequent construction of a predictive model that combines physical and data-driven approaches.

[0074] S202: Based on the operation time window of each process in the workshop production schedule and the equipment energy consumption benchmark corresponding to the order bill of materials, construct a pre-scheduled production feature sequence that is aligned with the historical power load data in terms of time resolution.

[0075] For example, this step aims to achieve data mode conversion from discrete business actions to continuous energy consumption waveforms. The system projects production scheduling information, originally centered on work order events, onto a unified time axis in conjunction with the physical power attributes of the equipment. This transforms unstructured production plans into numerical sequences that are isomorphic and frequency-similar to historical power load data, resolving the mismatch between business data and energy consumption data in terms of time granularity and data structure, and providing a standardized feature foundation for subsequent model input.

[0076] In one possible implementation, based on the operation time windows of each process in the workshop production schedule and the equipment energy consumption benchmarks corresponding to the order bill of materials, a pre-scheduled production feature sequence aligned with historical power load data in terms of time resolution is constructed. Specifically, this includes: parsing the material code information contained in the order bill of materials and querying the production equipment identifier associated with the material code information and the rated power parameters corresponding to the production equipment identifier in a preset equipment energy consumption database; parsing the workshop production schedule and extracting the start-stop time points and operating duration of the production equipment identifier corresponding to each process; generating a theoretical energy consumption curve based on the rated power parameters, start-stop time points, and operating duration; and resampling and interpolating the theoretical energy consumption curve according to the time sampling frequency of historical power load data to generate the pre-scheduled production feature sequence.

[0077] In this embodiment, the pre-scheduled production characteristic sequence refers to the expected energy consumption characteristic data continuously distributed over time, derived from future production plans. It is used to quantify the contribution of future production activities to the electricity load under ideal conditions. For example, this sequence is a set of time series vectors with the same time sampling frequency (e.g., one data point every 15 minutes) as historical electricity load data. Each value represents the theoretical power superposition value calculated based on the current production schedule, reflecting the expected energy consumption intensity of the production line at a specific point in time.

[0078] Specifically, the system performs deep analysis of the order bill of materials, extracting all material codes and using them as index keys to perform correlation searches in a pre-built equipment energy consumption database. This determines the specific production equipment identifier required to process each material code and obtains the rated power parameters corresponding to that equipment identifier (which may include standard energy consumption data such as equipment startup power, operating power, and standby power). Next, the system analyzes the workshop production schedule, extracting the specific operational arrangements for each production equipment identifier in the future time period at the process level, clarifying its precise start-up and shutdown times and operating duration.

[0079] Furthermore, the obtained rated power parameters are mapped onto the corresponding time axis. It is important to note that, to more closely resemble the real physical production environment, the theoretical energy consumption curve generated in this embodiment is not a simple "0-1" square wave, but a multi-state waveform constructed based on a refined equipment operating power spectrum. In specific implementation, the system divides the duration of each process into four stages: startup ramp-up, stable operation, idling standby, and shutdown cooling. For the startup ramp-up stage, an instantaneous peak power higher than the rated power is assigned based on the startup inrush current characteristics recorded in the database. For the stable operation stage, the rated power is weighted and corrected based on the material properties (such as hardness and material composition) or processing complexity coefficient in the order bill of materials (e.g., the power coefficient for processing high-hardness alloy steel is set to 1.2, and the coefficient for processing aluminum alloy is 0.8) to reflect the energy consumption differences under different loads. For the idling standby stage (such as the gap between processes), a lower standby power is assigned. Through the above processing, a discontinuous theoretical energy consumption curve reflecting the intermittent operation and load fluctuation characteristics of the equipment is generated. In order to enable the theoretical data to be jointly analyzed with the actual load data, the theoretical energy consumption curve is resampled and interpolated according to the predetermined time sampling frequency of historical power load data (e.g., 15 min / point) (e.g., using zero-order hold or linear interpolation method). This transforms discrete equipment start-up and shutdown events into continuous numerical signals that are fully aligned in time resolution, thereby generating the final pre-scheduled production characteristic sequence.

[0080] S203: Perform multi-dimensional feature splicing on historical power load data and pre-scheduled production characteristic sequences to generate a time-series input tensor.

[0081] For example, this step primarily involves the deep fusion and feature engineering construction of multi-source heterogeneous data. The system aligns and parallelizes electricity load data, representing historical energy consumption inertia, and pre-scheduled production sequences, representing future production intentions, on a unified time base. Through this multi-dimensional feature stitching, the originally scattered and independent univariate time series are integrated into a high-dimensional data structure containing multivariate coupling relationships. This provides a standardized data foundation for deep learning models that includes both historical trend information and future business increment information, ensuring that the model can comprehensively learn the changing patterns of electricity load from multiple perspectives.

[0082] In one possible implementation, historical power load data and pre-scheduled production feature sequences are multi-dimensionally concatenated to generate a time-series input tensor. Specifically, this includes: detecting and removing outliers from historical power load data, and filling the resulting data gaps using Lagrange interpolation; mapping the filled historical power load data and pre-scheduled production feature sequences to a preset numerical range using a max-min normalization algorithm; extracting calendar feature data corresponding to the target prediction period, where the calendar feature data includes at least weekday identifiers, holiday identifiers, and seasonal index identifiers; and concatenating the normalized historical power load data, the normalized pre-scheduled production feature sequences, and the calendar feature data in parallel along the feature dimension to construct a three-dimensional time-series input tensor.

[0083] In this embodiment, the time-series input tensor refers to a preprocessed and structured multidimensional data array used as a standard input interface for deep learning models (especially recurrent neural networks) to carry composite feature information containing time dependencies. Exemplarily, this time-series input tensor is typically constructed as a three-dimensional structure (e.g., number of samples × time step × feature dimension), where the feature dimension integrates the physical inertia of historical load, the business intent of future production scheduling, and the temporal attributes of the external environment, comprehensively reflecting the diverse factors affecting changes in power load.

[0084] Specifically, the system performs rigorous data cleaning on the collected raw historical power load data. Statistical methods (such as the 3σ principle or box plots) are used to identify and remove noise or erroneous readings that significantly deviate from the normal fluctuation range. For the resulting data gaps, Lagrange interpolation is applied, using valid data points before and after the gap to construct a polynomial function for fitting calculations, thereby accurately filling the gaps to ensure the continuity of the time series. To eliminate dimensional differences between different types of data and accelerate model convergence, a max-min normalization algorithm is used to calculate the extreme values ​​of historical power load data and pre-scheduled production characteristic sequences, linearly mapping them to a preset numerical range (usually [0, 1]). Simultaneously, the impact of time factors on load is explored, extracting calendar feature data corresponding to the target prediction period, and using one-hot encoding or embedding techniques to transform discrete date information into high-dimensional numerical vectors. Specifically, for weekday identifiers, a binary vector of length 7 is constructed (e.g., [1, 0, 0, 0, 0, 0, 0] represents Monday); for holiday identifiers, a binary variable (0 or 1) is constructed; for seasonal indexes, sine and cosine functions are used to map month information into continuous periodic feature values ​​(e.g., ...). To maintain the continuity of the time period, a tensor concatenation operation is performed. Assuming the feature dimension of historical power load data is D1, the feature dimension of the pre-scheduled production feature sequence is D2, the dimension of calendar feature data is D3, and the time step is T, the system stacks these matrices along the feature channel dimension to construct a three-dimensional time-series input tensor of shape (Batch_Size, T, D1+D2+D3). Batch_Size represents the number of independent samples processed in parallel during a single model computation (training or inference). This operation ensures that at each time step t with a time step of T, the model can simultaneously read the load value, production schedule energy consumption, and time attributes at that moment.

[0085] S204: Input the time-series input tensor into the pre-trained recurrent neural network prediction model to obtain the initial power load prediction value for the target prediction period.

[0086] For example, this step leverages the nonlinear fitting and time-series extrapolation capabilities of deep learning models to map multi-source heterogeneous data to electricity load values. The system uses tensor data carrying historical energy consumption patterns and future production task characteristics as driving input to drive a pre-trained recurrent neural network model for forward propagation calculations. During this process, the model makes a preliminary prediction of the load trend for the target period based on the learned "order-production scheduling-energy consumption" association rules, generating an initial electricity load prediction curve based on the ideal production scheduling execution state. This curve will serve as the basis for subsequent dynamic correction based on real-time production progress.

[0087] In one possible implementation, the time-series input tensor is input into a pre-trained recurrent neural network prediction model to obtain the initial power load prediction value for the target prediction period. Specifically, this includes: inputting the time-series input tensor into the input layer of the recurrent neural network prediction model, converting the time-series input tensor into a hidden layer feature vector through the input layer; processing the hidden layer feature vector through a long short-term memory network unit in the recurrent neural network prediction model to output a hidden state vector; and inputting the hidden state vector into a fully connected layer of the recurrent neural network prediction model for linear transformation to obtain the initial power load prediction value corresponding to the time step of the target prediction period.

[0088] In this embodiment, the recurrent neural network prediction model refers to a nonlinear mathematical model based on a deep learning algorithm architecture, specifically designed for processing and predicting time series data. It is used to capture long-term dependencies and complex temporal evolution patterns in data sequences. For example, this model mainly consists of a cascaded input layer, hidden layers (containing long short-term memory network units), and an output layer (fully connected layer). Through recursive calculations over time steps, it can effectively memorize historical information and predict future trends in power load changes.

[0089] Specifically, the system feeds the constructed time-series input tensor into the input layer of the recurrent neural network prediction model. Using the weight matrix of the input layer, it performs preliminary feature extraction and spatial mapping on the high-dimensional tensor data, converting it into a hidden layer feature vector suitable for internal network computation. Subsequently, this hidden layer feature vector is passed as input to the core component of the model—the Long Short-Term Memory (LSTM) network. The LSTM unit utilizes its internal gating mechanisms (including forget gate, input gate, and output gate) to perform temporal recursive processing on the feature vector, preserving key historical information while filtering redundant noise. This updates the memory unit state and outputs a hidden state vector carrying deep temporal dependency information. Finally, this hidden state vector is input to the fully connected layer of the recurrent neural network prediction model. Through linear transformation and dimensionality reduction operations using the weight matrix and bias terms of the fully connected layer, the high-dimensional hidden state is mapped back to the numerical space of the target physical quantity, thereby calculating the initial power load prediction value corresponding one-to-one with the time step of the target prediction period.

[0090] In one possible implementation, before inputting the time-series input tensor into the pre-trained recurrent neural network prediction model, the method further includes a training step for constructing the recurrent neural network prediction model. The training step specifically includes: acquiring historical measured load data, historical actual production schedule records, and historical calendar feature data within a preset historical time period; performing feature processing on the historical actual production schedule records to construct a historical production feature sequence aligned with the historical measured load data in the time dimension; concatenating the historical measured load data, historical production feature sequence, and historical calendar feature data in the feature dimension to generate a historical training sample set; inputting the historical training sample set into the recurrent neural network prediction model to be trained, calculating the loss function value between the model output and the historical measured load data, and updating the weight parameters of the recurrent neural network prediction model to be trained using an optimization algorithm until the loss function value meets a preset convergence condition, thereby obtaining the pre-trained recurrent neural network prediction model.

[0091] In this embodiment, the historical training sample set refers to a set of standardized, time-aligned, and supervised historical data, which serves as the learning material for a deep learning model, driving the model to capture the nonlinear mapping relationship between input variables and output results through iterative computation. For example, this sample set contains a large number of sample pairs, each consisting of input features (such as load at a past moment, production intensity at that time, and date type at that time) and a corresponding target ground truth (such as the actual electricity load value at that moment), aiming to enable the model to establish a nonlinear mapping relationship between specific production and time features and the true value of electricity load.

[0092] Specifically, the system needs to backtrack and extract raw data resources from the enterprise's historical database within a preset historical time period (e.g., the past year). This includes historical measured load data recording actual energy consumption, historical actual production schedule records reflecting the actual operating status of equipment (i.e., execution logs in the MES system), and corresponding historical calendar feature data. The unstructured historical actual production schedule records are then characterized by parsing the equipment start-up and shutdown times and operating parameters in historical work orders. This is transformed into a numerical historical production feature sequence that is strictly aligned with the historical measured load data in the time dimension, thereby quantifying the theoretical energy intensity of historical production behavior. In terms of feature dimension, the normalized historical measured load data, the generated historical production feature sequence, and the historical calendar feature data are concatenated in parallel to construct a historical training sample set containing a large number of input-label pairs. In the supervised learning phase of the model, the historical training sample set is fed into the recurrent neural network prediction model to be trained in batches, and forward propagation calculations are performed to obtain the prediction results.

[0093] Furthermore, to ensure the recurrent neural network prediction model has good generalization ability when dealing with non-stationary power load sequences and to avoid overfitting, this embodiment employs the following optimization strategies during training: Network structure optimization: A random deactivation layer is introduced between the LSTM layer and the fully connected layer. During training, neuron connections are randomly disconnected with a preset probability (e.g., 0.2), forcing the network to learn more robust distribution characteristics and preventing the model from rote memorization of specific historical data. Loss function design: Considering the sensitivity of industrial load forecasting to peak values, this embodiment uses a combined loss function to calculate the deviation between the model output and historical measured load data. For example, the mean squared error (MSE) and mean absolute percentage error (MAPE) are weighted and summed. For example: Loss function ,in The balancing coefficient is exemplarily set to 0.7. MSE is used to penalize large prediction biases to guide the model towards rapid convergence, while MAPE is used to balance relative errors under low load conditions, ensuring prediction accuracy during low-load periods (such as nighttime). Dynamic learning rate adjustment is implemented by introducing a learning rate decay mechanism when updating weight parameters through backpropagation using optimization algorithms such as Adaptive Moment Estimation (Adam). Initially, a large learning rate (e.g., 0.001) is used to rapidly descent the gradient. When the validation set stops decreasing, the learning rate is proportionally reduced (e.g., decayed to 0.1 times) to help the model perform fine-tuning searches near local minima. This continues until the loss function value meets preset convergence conditions (e.g., error less than a threshold or the number of iterations reaches a certain threshold), ultimately resulting in a pre-trained recurrent neural network prediction model with converged parameters and high-precision prediction capabilities.

[0094] In one possible implementation, the historical actual production scheduling records are characterized to construct a historical production feature sequence aligned with the historical measured load data in the time dimension. Specifically, this includes: removing abnormal record entries in the historical actual production scheduling records that are not aligned with the timestamps of the historical measured load data; parsing the production equipment operation data contained in the historical actual production scheduling records, and generating a historical energy consumption curve based on the production equipment operation data and preset rated power parameters; resampling and interpolating the historical energy consumption curve according to the time sampling frequency of the historical measured load data to generate a historical production feature sequence; and using a maximum-minimum normalization algorithm to map the historical production feature sequence to a preset numerical range consistent with the historical measured load data.

[0095] In this embodiment, the historical production feature sequence refers to a digital feature vector that strictly corresponds to the power load record in the time dimension, generated based on past real production records after cleaning, transformation, and mathematical processing. It is used to quantitatively characterize the theoretical contribution intensity of production activities to the power load within a historical time period. For example, this sequence is a set of continuous time series data, where each data point corresponds to a sampling moment of historical measured load data, and the numerical value reflects the theoretical total power superposition state of all operating equipment at that moment.

[0096] Specifically, the system performs quality screening on the acquired original historical production scheduling records, comparing the timestamp information in the records with the effective time span of the historical measured load data. Records with missing timestamps, incorrect formats, or exceeding the coverage of load data are identified as abnormal data and removed to ensure the consistency of the data source in the time base for subsequent processing. The cleaned historical production scheduling records are analyzed in depth to extract key status data about the actual operation of production equipment (such as actual start-up time, completion time, and equipment ID). Based on the material attributes and equipment operating stages (such as startup, stable, and standby) corresponding to historical work orders, preset rated power parameters and corresponding power correction coefficients are indexed to construct a discontinuous historical energy consumption curve reflecting the actual start-up and shutdown status of equipment and details of load fluctuations. To address the mismatch between the event-driven characteristics of the production scheduling records and the time sampling characteristics of the load data, the historical energy consumption curve is resampled and interpolated according to a predetermined time sampling frequency of the historical measured load data (e.g., once every 15 minutes), transforming the step-change power curve into a discrete but time-aligned numerical sequence. To eliminate dimensional differences and accelerate model training convergence, the maximum and minimum values ​​in the sequence are calculated using the maximum-minimum normalization algorithm. All values ​​in the sequence are then linearly mapped to a preset value range (e.g., [0, 1]) that is consistent with historical measured load data, thereby generating a historical production feature sequence that can be used for model training.

[0097] S205: Real-time monitoring of the actual production progress of the manufacturing execution system, and calculation of the time deviation between the actual production progress and the workshop production schedule.

[0098] In this embodiment, the time deviation value refers to a numerical indicator that quantifies the degree of lag or advancement in the actual execution of production compared to the predetermined plan in the time dimension. It is used as the basis for determining whether to trigger dynamic correction of the power load forecast results. For example, the deviation value is usually measured in minutes or hours. A positive value may indicate production lag, and a negative value may indicate production advancement. This value can accurately reflect the deviation in production rhythm caused by equipment failure, raw material delays, or efficiency fluctuations.

[0099] Specifically, by establishing a real-time communication interface with the MES (such as an MQTT message queue or API polling), the system continuously acquires actual production progress data from the workshop floor. This data includes the status of the currently processing steps, real-time equipment operation logs, and work order reporting records. The actual completion time or start time of the current step is then compared in real-time with the planned time node of the corresponding step in the pre-stored workshop production schedule. The difference between the actual time and the planned time is calculated (e.g., actual completion time minus planned completion time) to quantify the deviation of the current production pace from the original plan, thus calculating the time deviation value between the actual production progress and the workshop production schedule.

[0100] It should be noted that, in calculating time deviation values ​​and determining whether corrections are triggered, to avoid frequent model recalculations due to false deviations caused by data transmission delays in the MES system or human error in reporting work, the system employs a sliding time window weighted decision mechanism: deviation smoothing. The system does not rely on the deviation value at a single moment, but maintains a short-period deviation observation window (e.g., 5 sampling points within the past 30 minutes). An exponentially weighted moving average of the time deviation values ​​within this window is calculated, giving higher weight to the latest observation while retaining the smoothing effect of historical trends. Only when the smoothed deviation value exceeds a threshold is it considered a valid deviation. Dynamic threshold setting: The preset adjustment threshold is not fixed but is related to the current load baseline. During peak electricity load periods (high electricity prices, high demand), a smaller adjustment threshold is set (e.g., correction is triggered when the deviation exceeds 15 minutes) to ensure the model's sensitivity to load changes; during off-peak load periods, a larger adjustment threshold is set (e.g., correction is triggered when the deviation exceeds 45 minutes) to reduce the waste of computational resources.

[0101] S206: When the time deviation exceeds the preset adjustment threshold, the pre-scheduled production feature sequence is corrected based on the actual production progress, and the corrected pre-scheduled production feature sequence is re-inputted into the recurrent neural network prediction model to output the corrected power load prediction value.

[0102] For example, this step constructs a closed-loop power load forecasting correction mechanism based on production execution feedback. In actual industrial production, influenced by dynamic factors such as equipment failure and material flow, the actual production progress often deviates from the predetermined workshop production schedule, resulting in a phase misalignment between the initial power load forecast generated based on the static production schedule and the actual energy consumption. This step introduces a real-time deviation judgment mechanism. When a significant progress deviation is detected, the input data is dynamically refreshed in response to the deviation, adjusting the feature sequence originally based on the initial production schedule to a feature sequence that conforms to the current actual production state, and driving the recurrent neural network to perform secondary inference. This mechanism ensures that the prediction model can adapt to real-time changes in the production rhythm, thereby outputting corrected prediction values ​​that are synchronized with the physical production state in a cover-up update manner, minimizing prediction errors caused by production execution deviations.

[0103] In one possible implementation, when the time deviation value exceeds a preset adjustment threshold, the pre-scheduled production feature sequence is corrected based on the actual production progress. Specifically, this includes: determining the lag or advance time of the actual production progress relative to the workshop production plan based on the positive or negative attribute of the time deviation value; locating the feature data segment corresponding to the unexecuted process after the current time in the pre-scheduled production feature sequence; shifting the feature data segment along the time axis according to the lag or advance time to generate a shifted intermediate feature sequence; filling the empty time window caused by the shift in the intermediate feature sequence with zero values ​​or filling it with the mean value based on the feature values ​​of adjacent time steps to obtain the corrected pre-scheduled production feature sequence.

[0104] In this embodiment, the corrected pre-scheduled production characteristic sequence refers to new time series data generated after dynamically calibrating the original planned data based on real-time feedback from the production site. It is used to represent the expected future production energy consumption after correcting for time phase deviations. For example, the sequence retains the energy consumption waveform characteristics (i.e., waveform shape and amplitude) of each process in the original plan, but has been shifted as a whole on the time axis, thereby more accurately reflecting the change in load occurrence time due to delays or advances in the construction period.

[0105] Specifically, the system needs to analyze the sign and value of the calculated time deviation value, and qualitatively determine the current production status based on the positive or negative attribute of the deviation value: if the deviation value is positive, it is usually determined that the actual production progress is behind the plan, and is identified as a lag time; if the deviation value is negative, it is determined that the actual production progress is ahead of the plan, and is identified as an advance time. Taking the current moment as the dividing point, the system searches and extracts from the originally constructed pre-scheduled production feature sequence to locate the waveform data corresponding to those processes after the current moment that have not yet been executed, i.e., feature data segments. Based on the determined lag time or advance time, a rigid transformation operation is performed on the time axis direction of the feature data segment. For example, when lagging, the segment is shifted to the right in the future time direction, and when ahead, it is shifted to the left in the past time direction, thereby generating a shifted intermediate feature sequence with changed position. For gaps in the sequence caused by translation operations (such as gaps immediately following the current time due to right shift, or gaps at the end of the sequence due to left shift), a completion strategy is selected based on the actual process logic. For example, zero values ​​are filled if the equipment is considered to be in standby or idling state, or average values ​​are filled based on the characteristic trends of adjacent time steps, thereby ensuring the integrity and continuity of the sequence and obtaining the final corrected pre-scheduled production feature sequence.

[0106] This embodiment also discloses an electronic device, as shown in the reference. Figure 3 The electronic device may include: at least one processor 301, at least one communication bus 302, user interface 303, network interface 304, and at least one memory 305.

[0107] The communication bus 302 is used to enable communication between these components.

[0108] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0109] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0110] The processor 301 may include one or more processing cores. The processor 301 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 305, and by calling data stored in memory 305. Optionally, the processor 301 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 301 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 301 and may be implemented as a separate chip.

[0111] The memory 305 may include random access memory (RAM) or read-only memory. Optionally, the memory 305 may include a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 305 may also be at least one storage device located remotely from the aforementioned processor 301. Figure 3 As shown, the memory 305, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program based on an order-driven power load forecasting method.

[0112] exist Figure 3In the electronic device shown, the user interface 303 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 301 can be used to call an application program stored in the memory 305 that is a power load forecasting method based on an order. When executed by one or more processors 301, the electronic device executes one or more methods as described in the above embodiments.

[0113] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0114] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0115] In some embodiments of this application, a computer-readable storage medium is provided, including instructions that, when executed on an electronic device, cause the electronic device to perform an order-driven power load forecasting method according to an embodiment of this application.

[0116] In some embodiments of this application, a computer program product is also provided, which, when run on an electronic device, causes the electronic device to execute an order-driven power load forecasting method according to an embodiment of this application.

[0117] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0118] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0119] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0120] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 305 and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory 305 includes various media capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0121] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the disclosure in this specification. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are considered exemplary only, and the scope of this application is defined by the claims.

Claims

1. An order-driven power load forecasting method, characterized in that, The method includes: Historical power load data prior to the target forecast period and production business data corresponding to the target forecast period are obtained. The production business data includes order bill of materials and workshop scheduling plans from the manufacturing execution system. Based on the operation time window of each process in the workshop production schedule and the equipment energy consumption benchmark corresponding to the order bill of materials, a pre-scheduled production feature sequence aligned with the historical power load data in terms of time resolution is constructed. The historical power load data and the pre-scheduled production feature sequence are spliced ​​together using multi-dimensional features to generate a time-series input tensor. The time-series input tensor is input into a pre-trained recurrent neural network prediction model to obtain the initial power load prediction value for the target prediction period. Real-time monitoring of the actual production progress of the manufacturing execution system, and calculation of the time deviation between the actual production progress and the workshop production schedule; When the time deviation exceeds the preset adjustment threshold, the pre-scheduled production feature sequence is corrected based on the actual production progress, and the corrected pre-scheduled production feature sequence is re-inputted into the recurrent neural network prediction model to output the corrected power load prediction value.

2. The method according to claim 1, characterized in that, The step of constructing a pre-scheduled production feature sequence aligned with the historical power load data in terms of time resolution, based on the operation time windows of each process in the workshop production schedule and the equipment energy consumption benchmarks corresponding to the order bill of materials, specifically includes: The material code information contained in the order bill of materials is parsed, and the production equipment identifier associated with the material code information and the rated power parameter corresponding to the production equipment identifier are queried in the preset equipment energy consumption database; Analyze the workshop production schedule and extract the start and stop times and operating duration of the production equipment corresponding to each process; A theoretical energy consumption curve is generated based on the rated power parameters, the start-stop time points, and the operating duration. The theoretical energy consumption curve is resampled and interpolated according to the time sampling frequency of the historical power load data to generate the pre-scheduled production characteristic sequence.

3. The method according to claim 1, characterized in that, The step of concatenating the historical power load data and the pre-scheduled production feature sequence with multi-dimensional features to generate a time-series input tensor specifically includes: Outlier detection and removal are performed on the historical power load data, and the resulting data gaps are filled using Lagrange interpolation. The max-min normalization algorithm is used to map the filled historical power load data and the pre-scheduled production feature sequence to a preset numerical range; Extract calendar feature data corresponding to the target prediction period, wherein the calendar feature data includes at least weekday identifiers, holiday identifiers, and seasonal index identifiers; The normalized historical power load data, the normalized pre-scheduled production feature sequence, and the calendar feature data are concatenated in parallel along the feature dimension to construct the time-series input tensor with a three-dimensional structure.

4. The method according to claim 3, characterized in that, The step of inputting the time-series input tensor into a pre-trained recurrent neural network prediction model to obtain the initial power load prediction value for the target prediction period specifically includes: The temporal input tensor is fed into the input layer of the recurrent neural network prediction model, and the input layer converts the temporal input tensor into a hidden layer feature vector. The hidden layer feature vector is processed through the long short-term memory network unit in the recurrent neural network prediction model to output the hidden state vector; The hidden state vector is input into the fully connected layer of the recurrent neural network prediction model for linear transformation to obtain the initial power load prediction value corresponding to the time step of the target prediction period.

5. The method according to claim 3, characterized in that, Before inputting the temporal input tensor into the pre-trained recurrent neural network prediction model, the method further includes a training step for constructing the recurrent neural network prediction model, the training step specifically including: Acquire historical measured load data, historical actual production scheduling records, and historical calendar feature data within a preset historical time period; The historical actual production scheduling records are characterized to construct a historical production feature sequence that is aligned with the historical measured load data in the time dimension; The historical measured load data, the historical production feature sequence, and the historical calendar feature data are concatenated along the feature dimension to generate a historical training sample set. The historical training sample set is fed into the recurrent neural network prediction model to be trained. The loss function value between the model output and the historical measured load data is calculated. The weight parameters of the recurrent neural network prediction model to be trained are updated using an optimization algorithm until the loss function value meets the preset convergence condition, thus obtaining the pre-trained recurrent neural network prediction model.

6. The method according to claim 5, characterized in that, The step of performing feature processing on the historical actual production scheduling records to construct a historical production feature sequence aligned with the historical measured load data in the time dimension specifically includes: Remove abnormal record entries from the historical actual production scheduling records that are not aligned with the timestamps of the historical measured load data; The historical actual production scheduling records contain production equipment operation data, and historical energy consumption curves are generated based on the production equipment operation data and preset rated power parameters. The historical energy consumption curve is resampled and interpolated according to the time sampling frequency of the historical measured load data to generate the historical production characteristic sequence. The historical production feature sequence is mapped to a preset numerical range consistent with the historical measured load data using the max-min normalization algorithm.

7. The method according to claim 1, characterized in that, When the time deviation value exceeds a preset adjustment threshold, the pre-scheduled production characteristic sequence is corrected based on the actual production progress, specifically including: The lag or advance time of the actual production progress relative to the workshop production schedule is determined based on the positive or negative attribute of the time deviation value. Locate the feature data segment corresponding to the unexecuted process after the current time in the pre-scheduled production feature sequence; Based on the lag duration or the advance duration, the feature data segment is shifted along the time axis to generate the shifted intermediate feature sequence. The missing time windows caused by the translation of the intermediate feature sequence are filled with zero values ​​or filled with the mean value based on the feature values ​​of adjacent time steps to obtain the corrected pre-schedule production feature sequence.

8. An electronic device, characterized in that, Including processor and memory; The memory is used to store computer program code, the computer program code including computer instructions, and the processor invokes the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed on the electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on an electronic device, it causes the electronic device to perform the method as described in any one of claims 1-7.

Images