Power customer demand prediction method and system based on deep learning

Abstract

The invention discloses a power customer demand prediction method and system based on deep learning, and relates to the technical field of power system operation and power service intellectualization, and the method comprises the following steps: constructing a customer side equipment-behavior mutual identification decoding module, a power grid side constraint dynamic feedback module and a three-dimensional symbiosis prediction module; according to the invention, through the client-side equipment-behavior mutual identification decoding module, the demand essence is accurately identified, a three-dimensional scene gene label containing the equipment state, the behavior intention and the like is output, and demand misjudgment is avoided; through a power grid side constraint dynamic feedback module, a power grid bearing capacity dynamic curve and a constraint transfer suggestion are automatically generated, and the feasibility of a prediction result is guaranteed; a client side and power grid side double data link bidirectional interaction closed loop is constructed, self-iteration reinforcement learning of a three-dimensional symbiotic prediction module is combined, an isolated analysis framework is broken, collaborative decision is realized to support electric power service digital transformation, and meanwhile, an equipment hidden fault early warning function is derived.

Description

Technical Field

[0001] This invention relates to the field of intelligent technology for power system operation and power services, specifically to a method and system for predicting power customer demand based on deep learning. Background Technology

[0002] Existing deep learning-based electricity customer demand forecasting technologies, due to their isolated analytical framework, separate the two core stages of customer-side data computation and grid-side feasibility verification. This results in forecasts that fail to simultaneously satisfy both the accuracy of demand identification and the feasibility of implementation, ultimately failing to support collaborative decision-making between the grid and customers. This has become a core obstacle hindering the intelligent upgrading of electricity services. This framework problem is specifically manifested in two key sub-dimensions, both stemming from the underlying defects of the isolated framework:

[0003] 1. Sub-dimension 1: The customer-side "scenario-demand" is isolated, and the prediction only stays at the numerical surface and cannot identify the essence of the demand:

[0004] Existing technologies equate customer demand with electricity load values, and only analyze historical electricity consumption and weather data on the customer side through deep learning models to output future power / electricity consumption. However, this severs the connection between electricity consumption scenarios and demand types: load is the result of demand, while scenario is the cause of demand. Predictions that are divorced from scenario cannot reach the essence of demand.

[0005] Sub-dimensional root cause: The isolated framework defines demand forecasting as a single data task on the customer side, failing to recognize that demand is a dynamic requirement driven by scenarios, resulting in forecasts that can only output numerical values ​​and cannot guide service implementation.

[0006] 2. Sub-dimensional 2: The "customer demand - power grid carrying capacity" model is isolated, resulting in infeasible predictions and ultimately a meaningless conclusion.

[0007] When predicting customer demand, existing technologies only output conclusions based on customer-side data (such as "a certain enterprise needs to increase capacity by 800kVA"), but do not simultaneously access real-time grid capacity data (such as transformer capacity margin and line load rate) to verify feasibility. Existing technologies require manual querying of the GIS system to confirm the capacity of power points during on-site inspections, which precisely reflects this deficiency.

[0008] Sub-dimensional root cause: The isolated framework treats demand forecasting and grid dispatch as independent links, failing to recognize that feasibility is a prerequisite for forecast value, resulting in the results being disconnected from grid operation and unable to support collaborative decision-making.

[0009] Both of the aforementioned sub-dimensions stem from the fundamental flaws of isolated analytical frameworks: the technology breaks down demand forecasting into a fragmented process of customer-side calculation → subsequent querying on the grid side, rather than a unified system of data collaboration → model collaboration. This flaw prevents forecasting from simultaneously identifying the essence of demand and ensuring its feasibility for implementation, ultimately failing to support the goal of intelligent and differentiated services across the entire business expansion process, becoming a core bottleneck in the digital transformation of power services. Therefore, this paper provides a deep learning-based method and system for power customer demand forecasting to overcome these problems. Summary of the Invention

[0010] The purpose of this invention is to provide a method and system for predicting electricity customer demand based on deep learning, so as to solve the problems mentioned in the background art.

[0011] To address the aforementioned technical problems, this invention provides a deep learning-based method for predicting electricity customer demand, comprising the following steps:

[0012] Construct a customer-side device-behavior mutual verification decoding module, a power grid-side constraint dynamic feedback module, and a three-dimensional symbiotic prediction module to establish a two-way interactive closed loop between the customer-side data link and the power grid-side data link;

[0013] The customer-side device-behavior mutual verification decoding module collects customer-side device operation data, electricity consumption behavior trajectory data, and scenario auxiliary data. After preprocessing, the device-behavior mutual verification decoding is performed, and a three-dimensional scenario gene label including device status, behavior intent, demand type, and scenario confidence is output.

[0014] The grid-side constraint dynamic feedback module collects real-time grid data and topology data, and uses an LSTM model incorporating topology weight factors to predict grid carrying capacity, generating dynamic curves of grid carrying capacity and constraint transfer suggestions.

[0015] A three-dimensional symbiotic space of equipment, behavior, and constraints is constructed through a three-dimensional symbiotic prediction module. Self-iterative prediction is performed using three-dimensional reward function reinforcement learning, and a four-dimensional prediction report is output, including the result of demand nature identification, power grid feasibility scheme, and equipment health.

[0016] Furthermore, the customer-side device-behavior mutual verification decoding module collects data through the customer-side device-load dual-mode acquisition terminal, which includes a current harmonic sensor, a temperature sensor, a vibration sensor, a load acquisition module, and a human-machine interaction sub-terminal.

[0017] Furthermore, the preprocessing employs a sliding window-feature preservation algorithm, with the following specific steps:

[0018] Set a 10-minute window size and a 5-minute step size. Use the average load value within each window as the base load segment and the portion where the absolute value of the difference between the load value and the average value within the window is ≥1kW as the fluctuation characteristic segment. Standardize the equipment operation data into an equipment state vector, where the current harmonic distortion rate is mapped to the 0-1 interval by the actual value / 20%, the winding temperature to the actual value / 125℃, and the vibration frequency to the actual value / 1000Hz.

[0019] Furthermore, the device-behavior mutual verification decoding includes a dynamic device status fingerprint database construction step, which adopts an improved K-means clustering algorithm that introduces operating condition weight factors. The algorithm's clustering categories are divided into four types: normal operation, aging and decay, overload start-up, and fault precursor. The initialization method is K-means++, and it converges when the cluster center change is ≤0.01 after 50 iterations. The fingerprint database is initially generated from the operating data of 100 devices of the same model. If new features appear in the device data that do not match the existing fingerprints, incremental clustering is automatically triggered and synchronized to all client-side terminals.

[0020] Furthermore, the construction of the LSTM model incorporating topology weight factors includes: input features including time features, power grid features, and topology weight factors; adding topology weight factors to the conventional LSTM input gate, with the input gate activation function being: the model predicts the dynamic curve of power grid carrying capacity for the next 1-72 hours, including transformer capacity margin and line load rate threshold.

[0021] Furthermore, the formula for calculating the three-dimensional reward function is as follows:

[0022] R=0.3×R1+0.4×R2+0.3×R3;

[0023] in:

[0024] R1 is the scenario confidence reward: if the confidence level is ≥90%, R1 = 1; otherwise, R1 = 0.

[0025] R2 is the load-bearing capacity matching bonus: if the demand value range falls within the load-bearing capacity threshold, R2 = 1; otherwise, R2 = 0.

[0026] R3 is the device health bonus: if the device health score is ≥80, R3 = 1; otherwise, R3 = 0.

[0027] The self-iterative prediction process iterates for 5-6 rounds, converging when the reward value is ≥0.9 for two consecutive rounds. In each round, the demand implementation plan is adjusted based on the matching between the demand value range and the carrying capacity threshold.

[0028] Furthermore, it also includes a latent equipment fault early warning step: when the equipment health score is <70 or the fault risk probability is ≥30%, an early warning is triggered; the equipment health score and fault risk probability are calculated by comparing the current equipment state vector with the cosine similarity of the fault simulation library; the early warning information is pushed to the power grid operation and maintenance system via SMS and customer-side touch screen, and includes the suggested equipment to be repaired, the fault type, and the repair suggestions within 3 days.

[0029] The deep learning-based power customer demand forecasting system includes a customer-side device-load dual-mode acquisition terminal, a grid-side edge-topology dual-mode acquisition terminal, a customer-side device-behavior mutual verification decoding module, a grid-side constraint dynamic feedback module, a three-dimensional symbiotic prediction module, and a result output and early warning module. The customer-side acquisition terminal communicates with the customer-side decoding module, and the grid-side acquisition terminal communicates with the grid-side feedback module via the MQTT protocol.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] 1. By using the customer-side device-behavior mutual verification decoding module, the limitations of existing technologies that only output power load values ​​are overcome. Instead, a three-dimensional scene gene label containing device status, behavioral intent, demand type, and scene confidence level is output. The dynamic device status fingerprint database is constructed based on an improved K-means clustering algorithm that incorporates operating condition weight factors, adapting to individual differences among different customer devices. The improved DTW algorithm incorporates device feature constraints to achieve bidirectional mutual verification between behavioral trajectories and fingerprints. Combined with fault simulation back-inference verification, it distinguishes between energy consumption anomalies caused by equipment faults (such as load fluctuations caused by bearing wear) and actual power demand (such as production capacity expansion), avoiding misjudgment of demand and ensuring the accuracy of demand essence identification.

[0032] 2. Through the grid-side constraint dynamic feedback module (integrating an LSTM model with topology weight factors and constraint advance feedback logic), the system can automatically collect real-time grid data (transformer capacity margin, line load rate) and topology data without requiring manual querying of the GIS system to confirm power source capacity, predicting the grid carrying capacity dynamic curve for the next 1-72 hours. When the predicted carrying capacity is lower than the threshold (e.g., transformer capacity margin ≤ 50kVA, line load rate ≥ 90%), it can automatically retrieve backup power sources, calculate transferable capacity, and generate constraint transfer suggestions, enabling customer needs (e.g., capacity increase values) to be directly matched with grid carrying capacity, avoiding prediction results becoming unfeasible conclusions and supporting the feasibility of demand implementation plans.

[0033] 3. By constructing a two-way interactive closed loop between the customer-side data link (equipment operation data → behavioral trajectory data → scenario gene tags) and the power grid-side data link (real-time power grid data → dynamic carrying capacity curve → constraint transfer suggestions), and combining the three-dimensional reward function reinforcement learning of the three-dimensional symbiotic prediction module (integrating scenario confidence, carrying capacity matching, and equipment health rewards), 5-6 rounds of self-iterative prediction are executed (convergence occurs when the reward value is ≥0.9 for two consecutive rounds). The implementation plan is adjusted in each round based on the matching between the demand value range and the carrying capacity threshold. This transforms the original fragmented process of customer-side calculation → subsequent power grid-side query into a unified system of data collaboration → model collaboration, ensuring that the prediction results simultaneously meet the requirements of accurate demand identification and feasibility for implementation. This supports intelligent and differentiated services throughout the entire business expansion process, breaking through the core bottleneck of the digital transformation of power services.

[0034] 4. The system utilizes a latent equipment fault early warning function. Based on the cosine similarity between the current equipment state vector and the fault simulation library, it calculates the equipment health score and fault risk probability. When the equipment health score is <70 or the fault risk probability is ≥30%, it automatically pushes warning information via SMS and customer-side touchscreen, and synchronizes it to the power grid operation and maintenance system. This allows for the early detection of latent equipment faults, preventing equipment failures from affecting the assessment of electricity demand. It also assists the power grid operation and maintenance department in accurately scheduling maintenance, reducing the risk of power outages for customers, and improving the timeliness and reliability of power services.

[0035] 5. Customer-side data preprocessing employs a sliding window-feature retention algorithm and standardizes equipment operation data to the 0-1 range, accurately extracting load fluctuation characteristics to provide high-quality data for subsequent decoding. The dual-mode acquisition terminals on the customer and grid sides transmit data in real time via the MQTT protocol, and modules exchange structured data through API interfaces to ensure the real-time nature of data acquisition, processing, and prediction. The incremental update mechanism of the dynamic equipment status fingerprint database adapts to changes in equipment operating status and individual differences in equipment among different customers, ensuring the long-term adaptability and accuracy of the prediction model. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the power customer demand forecasting method and system based on deep learning, as described in this invention.

[0037] Figure 2 This invention provides a four-dimensional prediction report diagram in the power customer demand prediction method and system based on deep learning.

[0038] Figure 3 This is a system architecture diagram of the deep learning-based electricity customer demand forecasting method and system of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figures 1-3 The present invention provides a technical solution:

[0041] See Figures 1-3 As shown, an embodiment of a deep learning-based electricity customer demand forecasting method and system is presented:

[0042] 1. Purpose:

[0043] By constructing a three-dimensional symbiosis of equipment, behavior, and constraints, the problem of isolated scenarios and demands, and disconnect between demands and grid carrying capacity is solved simultaneously. This enables accurate identification of the essence of demands and direct implementation of prediction results, while also generating functions for predicting the health of customer equipment and dynamically optimizing grid resources.

[0044] 2. Core Solution:

[0045] Employing a three-module collaborative approach + dual data links:

[0046] Three-module collaboration: The customer-side device-behavior mutual verification decoding module realizes the identification of the essence of demand, the grid-side constraint dynamic feedback module generates grid carrying capacity constraints, and the three-dimensional symbiotic prediction module completes the collaborative iteration of demand and constraints;

[0047] Dual data links: The customer-side data link (equipment operation data → behavior trajectory data → scenario gene tags) and the power grid-side data link (real-time power grid data → dynamic carrying capacity curve → constraint transfer suggestions) interact bidirectionally to form a closed loop of data acquisition-processing-prediction-feedback.

[0048] Specific steps:

[0049] Step 1: Client-side device - Behavior verification decoding (used to solve the problem of isolated scenarios and requirements):

[0050] This step collects data through the customer-side device-load dual-mode acquisition terminal, and the device-behavior mutual verification decoding module performs a three-step decoding process of fingerprint database construction, two-way mutual verification, and fault reverse inference, outputting a three-dimensional scene gene label.

[0051] 1.1 Hardware Deployment and Data Acquisition:

[0052] Hardware Selection: Deployment of client-side equipment - Dual-mode load acquisition terminal (hardware composition and parameters are as follows):

[0053]

[0054]

[0055] Data acquisition range and frequency: Equipment operation data: current harmonic distortion rate (1 time / second), winding temperature (1 time / minute), vibration frequency (1 time / second); Electricity consumption behavior trajectory data: load value (1 time / minute), duration of load surge / descent (triggered acquisition, recorded when load change ≥1kW), peak occurrence period (automatically marks daily load peak and duration); Scenario auxiliary data: manually input via touch screen, supporting 3 fixed options: adding equipment, equipment maintenance, and production adjustment, as well as custom text input (automatically synchronized to the mutual verification decoding module after input).

[0056] 1.2 Data Preprocessing:

[0057] The sliding window-feature preservation algorithm is used to process the load data. The specific steps are as follows:

[0058] Set the sliding window parameters: window size 10 minutes, step size 5 minutes;

[0059] Load data splitting: Calculate the average load value within each window as the base load segment; the portion where the absolute value of the difference between the load value within a window and the average value is ≥1kW is taken as the fluctuation characteristic segment (e.g., "a sudden load increase lasting 30 minutes" corresponds to the fluctuation characteristic segment of 6 consecutive windows).

[0060] Equipment data standardization: Converting equipment operation data into equipment state vectors.

[0061] V eq = [HD,T,F];

[0062] in:

[0063] HD (Harmonic Distortion Rate): Normalized to the 0-1 range (actual value / 20%, since the maximum sensor measurement value is 20%).

[0064] T (winding temperature): Standardized to the 0-1 range (actual value / 125℃, since the maximum measurement value of the sensor is 125℃).

[0065] F (vibration frequency): Normalized to the 0-1 range (actual value / 1000Hz, since the maximum measurement value of the sensor is 1000Hz).

[0066] 1.3 Device-Behavior Mutual Authentication Decoding:

[0067] The device-behavior mutual verification decoding module achieves decoding through the following three steps. The module hardware is an industrial-grade server (configuration: CPU Intel Xeon E3-1230v5, memory 16GB, hard drive 1TB SSD). The software is developed using Python 3.8 and relies on the TensorFlow 2.5 framework.

[0068] 1.3.1: Step 1: Construction of Dynamic Device Status Fingerprint Database:

[0069] Algorithm selection: Improved K-means clustering algorithm (introducing working condition weight factor);

[0070] Clustering parameter settings:

[0071] Number of clusters: 4 (normal operation, aging and decay, overload start-up, and early signs of failure);

[0072] Initialization method: K-means++ (to avoid local optima);

[0073] Number of iterations: 50 (convergence threshold: cluster center change ≤ 0.01);

[0074] Calculation of operating condition weighting factor:

[0075] W = 1 + 0.2 × S ，

[0076] Where S is the working condition identifier, S=1 for production working conditions and S=0 for non-production working conditions, that is, the cluster radius is reduced by 20% under production working conditions;

[0077] Fingerprint database dynamic update mechanism:

[0078] Initial fingerprint database: collect operating data from 100 devices of the same model (such as injection molding machines), and generate initial 4 types of fingerprints using the above algorithm (each type of fingerprint contains feature thresholds, as shown in Table 1);

[0079] Incremental update: The client-side terminal uploads device data in real time. If a device data has a new feature that does not match the existing fingerprint (such as a harmonic distortion rate of 7% + load ramp rate of 1kW / minute), incremental clustering is automatically triggered (only 1 new cluster center is added) to supplement the new fingerprint (such as half-overload start-up) and synchronized to the fingerprint database of all client-side terminals.

[0080] Table 1 Initial Device State Fingerprint Feature Threshold Table:

[0081]

[0082] 1.3.2: Step 2: Two-way verification between behavioral trajectory and fingerprint:

[0083] An improved Dynamic Time Warping (DTW) algorithm is adopted (introducing device characteristic constraint terms), with the following specific steps:

[0084] 1.3.2.1 Constructing the input sequence:

[0085] Behavioral trajectory sequence:

[0086] X = [x1, x2, ..., x n ];

[0087] x1 is the maximum value of the fluctuation characteristic segment of the i-th sliding window; n is the number of windows on the day, such as 288 windows for 24 hours;

[0088] Fingerprint feature sequence:

[0089] Y = [y1, y2, ..., y m ];

[0090] y j is the load surge rate threshold for a certain type of fingerprint; m is the number of fingerprint categories, i.e., 4 categories;

[0091] 1.3.2.2 Calculate DTW distance (basic similarity):

[0092]

[0093] in:

[0094] d(x i ,y j )=|x i -y j | represents the Euclidean distance, w i,j The path weights (satisfying dynamic regularization constraints);

[0095] 1.3.2.3 Equipment Characteristic Constraint Item C:

[0096]

[0097] Example: When matching fingerprints for overload startup, both HD ≥ 8% and F ≥ 50Hz must be met simultaneously; otherwise, C = ∞. ， DTW distance is invalid;

[0098] 1.3.2.4. Bidirectional matching determination:

[0099] Forward matching: Calculate DTW(X,Y)+C between X and each class Y, and take the fingerprint corresponding to the minimum value as the candidate fingerprint;

[0100] Reverse verification: Extract the device feature threshold corresponding to the candidate fingerprint, and verify whether the current device state vector meets the requirements (e.g., if the candidate fingerprint is "overload start", then verify that HD≥8% and F≥50Hz. If the verification passes, the matching fingerprint is determined; otherwise, rematching is performed).

[0101] 1.3.3, Step 3: Fault simulation and reverse verification:

[0102] Fault simulation library construction: Using a general power laboratory (configured with the same type of equipment and fault simulation device), 10 common equipment faults were simulated, and the fault characteristic-behavioral deviation mapping relationship was recorded (as shown in Table 2);

[0103] Table 2 Equipment Failure-Behavioral Deviation Mapping Table:

[0104]

[0105] Reverse verification steps:

[0106] If the behavior trajectory matches the fingerprint of a scenario such as production capacity expansion or equipment addition (e.g., the overload start fingerprint corresponds to the production capacity expansion scenario);

[0107] Extract the current device state vector and compare it with the fault features in the fault simulation library. If the similarity is ≥80% (calculated using cosine similarity);

[0108] Trigger scenario re-judgment: Abnormal energy consumption caused by output device failure, which is not a real energy expansion demand, and mark the failure type (such as "bearing wear").

[0109] 1.4 Output Results:

[0110] Through the above three steps, the 3D scene gene tags are output:

[0111]

[0112] in:

[0113] S eq (Equipment Status): Values ​​include normal operation, aging and degradation, overload start-up, and fault precursor.

[0114] I act (Behavioral Intent): Values ​​include production capacity expansion, new equipment addition, equipment maintenance, and malfunction / abnormality.

[0115] D req (Demand type): such as short-term capacity increase of 200kVA, equipment maintenance demand, no new demand (demand value is calculated by integrating the fluctuation characteristic segment, such as "total load increase for 30 minutes" corresponding to capacity increase demand);

[0116] Conf (Scenario Confidence): Value ranges from 0 to 100% (calculated by combining the DTW distance of two-way mutual verification and the similarity of fault inference; for example, the smaller the DTW distance and the lower the fault similarity, the higher the confidence).

[0117] Step 2: Dynamic feedback of grid-side constraints: (used to address the disconnect between demand and grid capacity)

[0118] This step involves collecting data through a dual-mode acquisition terminal on the grid side edge-topology, and generating a dynamic curve of grid carrying capacity plus constraint transfer suggestions by the constraint dynamic feedback module.

[0119] 2.1 Hardware Deployment and Data Acquisition:

[0120] Hardware selection: Deploy a grid-side edge-topology dual-mode data acquisition terminal (hardware components are as follows):

[0121]

[0122]

[0123] Data collection scope and frequency:

[0124] Real-time power grid data: Transformer capacity margin (1 time / second), line load rate (1 time / second);

[0125] Topology data: Pre-stored physical distance between customer access point and transformer (error ≤ 10 meters), and connection relationship between transformer and backup power supply (e.g., "#1 transformer backup power supply is #2 transformer");

[0126] Historical data: "Constraint conflict records" for the past 3 months (e.g., "2024-05-10 14:00 #1 transformer capacity expansion demand exceeds margin of 50kVA").

[0127] 2.2 Constrained Dynamic Evolution Model:

[0128] The constraint dynamic feedback module is deployed at the grid-side edge gateway and uses a topology weighted LSTM model to predict grid carrying capacity. Specific steps include:

[0129] 2.2.1 Input Feature Construction:

[0130] Time characteristics: hour, day of the week, whether it is a holiday (coded as 0-1 variables);

[0131] Power grid characteristics: Transformer capacity margin (average value over the past hour), line load rate (maximum value over the past hour);

[0132] Topological weighting factor W topo :

[0133]

[0134] in:

[0135] L is the actual distance (in meters) between the customer access point and the transformer. L = 500 meters (preset maximum service radius). That is, the closer the distance, the higher the weight (maximum 1.0, minimum 0.8).

[0136] 2.2.2 Model Structure Improvement:

[0137] Add W to the input gate of a conventional LSTM topo The input gate activation function is calculated as follows:

[0138] i t =σ(W i ·[h t-1 ,x t ]+b i +W topo ·x t );

[0139] in:

[0140] i t For input gate output, W i Let h be the weight matrix. t-1 Given the hidden state of the previous time step, x t Input features for the current time step, b i Here, σ is the bias term, and σ is the sigmoid function.

[0141] 2.2.3 Predicted Output:

[0142] Forecast duration: 1-72 hours from now (1-hour increments);

[0143] Output: Hourly "Dynamic Curve of Power Grid Capacity" (including transformer capacity margin and line load rate threshold).

[0144] 2.3 Constrained advance feedback logic:

[0145] When the predicted grid carrying capacity is below a threshold for a certain period (e.g., transformer capacity margin ≤ 50kVA, line load rate ≥ 90%), constraint transfer suggestions are automatically triggered for generation.

[0146] Retrieve backup power supply: Query the backup power supply of the current transformer in the pre-stored topology (e.g., #1 transformer is backed up as #2 transformer);

[0147] Calculate transferable capacity:

[0148] C trans =C spare -C spare_used ;

[0149] in:

[0150] C spare For the prediction margin of backup power, C spare_used The backup power supply has been allocated to meet the demand;

[0151] Generate suggestions: such as "It is recommended that a customer's capacity expansion needs be transferred to transformer #2, which can release 150kVA of capacity. The transfer period is from 22:00 to 6:00 the next day".

[0152] Data interaction: The "constraint transfer suggestion" is pushed to the "3D symbiotic prediction module" via the MQTT protocol, and simultaneously synchronized to the client-side device - behavior mutual verification decoding module.

[0153] Step 3: 3D co-occurrence prediction (for solving the frame isolation problem):

[0154] This step is executed by the 3D symbiotic prediction module (deployed on the power grid side cloud server, configured with: CPU Intel Xeon Gold 6248, memory 64GB, hard disk 2TB SSD), which achieves collaborative decision-making through "3D symbiotic space construction - self-iterative prediction - result output".

[0155] 3.1 Construction of Three-Dimensional Symbiotic Space:

[0156] Define a three-dimensional symbiotic space of device-behavior-constraint, with the following inputs and parameters for the three dimensions:

[0157]

[0158]

[0159] 3.2 Self-iterative co-occurrence prediction:

[0160] The reinforcement learning agent employs a three-dimensional reward function, with specific parameters and steps as follows:

[0161] 3.2.1 Agent initialization:

[0162] State space: All parameter combinations in the three-dimensional symbiotic space; Action space: The range of adjustment requirements (e.g., "increase capacity in two batches" or "transfer to backup power");

[0163] 3.2.2 Three-dimensional reward function R:

[0164] R=0.3×R1+0.4×R2+0.3×R3;

[0165] in:

[0166] R1 (Scenario Confidence Reward): If the confidence level is ≥90%, R1 = 1; otherwise, R1 = 0.

[0167] R2 (Bearing Capacity Matching Bonus): If the demand value range falls within the bearing capacity threshold, R2 = 1; otherwise, R2 = 0.

[0168] R3 (Device Health Bonus): If the device health score is ≥80, R3 = 1; otherwise, R3 = 0.

[0169] 3.2.3 Iterative Process:

[0170] Number of iterations: 5-6 rounds (convergence condition: reward value R ≥ 0.9 for 2 consecutive rounds);

[0171] The adjustment logic for each round is as follows: "The demand dimension shows that 200kVA of capacity needs to be added, and the constraint dimension shows that the current margin is 150kVA → the agent is adjusted to 'increase capacity in two batches, the first batch of 150kVA (matching the current margin), and the second batch of 50kVA (matching the backup power margin for the next day)'."

[0172] 3.3 Output Results and Warning Triggering:

[0173] 3.3.1 Four-dimensional prediction report (structured output, JSON format): See [link / reference] Figure 2 As shown.

[0174] 3.3.2 Early warning of latent equipment faults:

[0175] Triggering conditions: Equipment health score < 70 points or failure risk probability ≥ 30%;

[0176] Warning output: Push notifications via SMS (connected to the power grid SMS gateway) and customer-side touchscreens, such as "It is recommended to inspect a certain piece of equipment (e.g., 'Injection Molding Machine 1#'), fault type 'bearing wear', and recommended inspection time 'within 3 days'".

[0177] Data synchronization: Early warning information is synchronized to the power grid operation and maintenance system (interfacing with the marketing management system API).

[0178] 4. System Composition and Data Interaction:

[0179] 4.1 System Overall Architecture:

[0180] This system consists of 5 core modules, which interact with each other via the MQTT protocol and API interfaces. See the architecture section below. Figure 3 As shown.

[0181] 4.2 Core Module Hardware and Software Configuration:

[0182]

[0183]

[0184] 4.3 Data Interaction Protocols and Interfaces:

[0185] MQTT protocol: used for collecting real-time data transmission between the terminal and the decoding / feedback module. The topic is defined as follows:

[0186] Client side: / client / device / data / [Client ID] (Device operation data)

[0187] / client / behavior / data / [customer ID](behavior trajectory data);

[0188] Grid side: / grid / capacity / data / [transformer ID] (capacity data)

[0189] / grid / constraint / advice / [transformer ID](constraint transfer suggestion);

[0190] API interface: Used for structured data exchange between modules, as shown in the example below:

[0191] Device-Behavior Mutual Authentication Decoding Module → 3D Symbiosis Prediction Module: POST / api / v1 / scene / tag (Input: 3D scene gene tag, Output: Receive confirmation);

[0192] Constraint Dynamic Feedback Module → 3D Symbiotic Prediction Module: POST / api / v1 / grid / constraint (Input: Bearing capacity curve + transfer suggestion, Output: Receive confirmation);

[0193] 3D symbiotic prediction module → Result output module: GET / api / v1 / prediction / report / [work order ID] (Input: work order ID, Output: 4D prediction report).

[0194] Summarize:

[0195] Dynamic device status fingerprint database: Breaking through the limitations of existing fixed fingerprint databases, it solves the problem of individual differences in devices of different customers by updating incrementally with real-time customer data;

[0196] Improved DTW two-way mutual verification algorithm: Through device feature constraints, it realizes the causal chain verification of behavior trajectory → device status → scene intent;

[0197] Topology Weight-LSTM Model: Transforms the grid topology distance into LSTM input gate weights to achieve spatial-temporal co-prediction of carrying capacity;

[0198] Three-dimensional reward function reinforcement learning: integrates three-dimensional rewards of scene confidence, carrying capacity matching and equipment health to ensure that the prediction results simultaneously meet the requirements of essential recognition and implementation feasibility.

Claims

1. A deep learning-based method for predicting electricity customer demand, characterized in that, Includes the following steps: Construct a customer-side device-behavior mutual verification decoding module, a power grid-side constraint dynamic feedback module, and a three-dimensional symbiotic prediction module to establish a two-way interactive closed loop between the customer-side data link and the power grid-side data link; The customer-side device-behavior mutual verification decoding module collects customer-side device operation data, electricity consumption behavior trajectory data, and scenario auxiliary data. After preprocessing, the device-behavior mutual verification decoding is performed, and a three-dimensional scenario gene label including device status, behavior intent, demand type, and scenario confidence is output. The grid-side constraint dynamic feedback module collects real-time grid data and topology data, and uses an LSTM model incorporating topology weight factors to predict grid carrying capacity, generating dynamic curves of grid carrying capacity and constraint transfer suggestions. A three-dimensional symbiotic space of equipment, behavior, and constraints is constructed through a three-dimensional symbiotic prediction module. Self-iterative prediction is performed using three-dimensional reward function reinforcement learning, and a four-dimensional prediction report is output, including the result of demand nature identification, power grid feasibility scheme, and equipment health.

2. The deep learning-based electricity customer demand forecasting method as described in claim 1, characterized in that: The customer-side device-behavior mutual verification decoding module collects data through the customer-side device-load dual-mode acquisition terminal, which includes a current harmonic sensor, a temperature sensor, a vibration sensor, a load acquisition module, and a human-machine interaction sub-terminal.

3. The deep learning-based electricity customer demand forecasting method as described in claim 1, characterized in that: The preprocessing uses a sliding window feature preservation algorithm, and the specific steps are as follows: Set a 10-minute window size and a 5-minute step size. Use the average load value within each window as the base load segment and the portion where the absolute value of the difference between the load value and the average value within the window is ≥1kW as the fluctuation characteristic segment. Standardize the equipment operation data into an equipment state vector, where the current harmonic distortion rate is mapped to the 0-1 interval by the actual value / 20%, the winding temperature to the actual value / 125℃, and the vibration frequency to the actual value / 1000Hz.

4. The deep learning-based electricity customer demand forecasting method as described in claim 1, characterized in that: The device-behavior mutual verification decoding includes the construction of a dynamic device status fingerprint database, which adopts an improved K-means clustering algorithm that introduces operating condition weight factors. The algorithm's clustering categories are divided into four types: normal operation, aging and decay, overload start-up, and fault precursor. The initialization method is K-means++, and it converges when the cluster center change is ≤0.01 after 50 iterations. The fingerprint database is initially generated from the operating data of 100 devices of the same model. If new features appear in the device data that do not match the existing fingerprints, incremental clustering is automatically triggered and synchronized to all client-side terminals.

5. The deep learning-based electricity customer demand forecasting method as described in claim 1, characterized in that: The construction of the LSTM model incorporating topology weighting factors includes: input features including time features, power grid features, and topology weighting factors; adding topology weighting factors to the conventional LSTM input gate, with the input gate activation function being: the model predicts the dynamic curve of power grid carrying capacity for the next 1-72 hours, including transformer capacity margin and line load rate threshold.

6. The deep learning-based electricity customer demand forecasting method as described in claim 1, characterized in that: The formula for calculating the three-dimensional reward function is: R=0.3×R1+0.4×R2+0.3×R3; in: R1 is the scenario confidence reward: if the confidence level is ≥90%, R1 = 1; otherwise, R1 = 0. R2 is the load-bearing capacity matching bonus: if the demand value range falls within the load-bearing capacity threshold, R2 = 1; otherwise, R2 = 0. R3 is the device health bonus: if the device health score is ≥80, R3 = 1; otherwise, R3 = 0. The self-iterative prediction process iterates for 5-6 rounds, converging when the reward value is ≥0.9 for two consecutive rounds. In each round, the demand implementation plan is adjusted based on the matching between the demand value range and the carrying capacity threshold.

7. The deep learning-based electricity customer demand forecasting method as described in claim 1, characterized in that: It also includes a hidden fault warning procedure: when the equipment health score is <70 or the fault risk probability is ≥30%, an early warning is triggered; the equipment health score and fault risk probability are calculated by comparing the current equipment state vector with the cosine similarity of the fault simulation library; the early warning information is pushed to the power grid operation and maintenance system via SMS and customer-side touch screen, and includes the suggested equipment to be repaired, the fault type and the repair suggestions within 3 days.

8. A deep learning-based power customer demand forecasting system, characterized in that: It includes a customer-side device-load dual-mode acquisition terminal, a grid-side edge-topology dual-mode acquisition terminal, a customer-side device-behavior mutual verification decoding module, a grid-side constraint dynamic feedback module, a three-dimensional symbiotic prediction module, and a result output and early warning module; the customer-side acquisition terminal and the customer-side decoding module, and the grid-side acquisition terminal and the grid-side feedback module communicate via the MQTT protocol.

Images