Pumped storage power station construction site electrical load accurate unit prediction system and method
By using multi-source data acquisition and a spatiotemporal collaborative multi-scale residual cyclic converter model, the accuracy and adaptability issues of power load forecasting at pumped storage power station construction sites were resolved. This resulted in high-precision load forecasting and intelligent safety early warning, improving the timeliness of load forecasting and carbon emission assessment at construction sites.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-17
- Publication Date
- 2026-03-13
AI Technical Summary
Existing methods for predicting electricity load at pumped storage power station construction sites are inadequate in terms of accuracy, timeliness, and adaptability. They also lack intelligent safety early warning and carbon emission assessment, and are ill-equipped to handle multi-source data processing and dynamic topological relationships in complex construction environments.
The system employs a multi-source data acquisition module, a spatiotemporal feature matrix construction module, and a load prediction module. It combines a spatiotemporal collaborative multi-scale residual cyclic transformer model, separates steady-state load from noise using wavelet basis functions and gated recursive units, processes spatiotemporal features using graph attention networks and ProbSparse multi-head attention mechanisms, outputs unit-level load curves, and provides hierarchical early warning through an early warning and scheduling decision module.
It significantly improves the accuracy and adaptability of unit-level load forecasting, realizes real-time safety early warning and equipment scheduling, and enhances the timeliness of load forecasting and the level of intelligence in carbon emission assessment at construction sites.
Smart Images
Figure CN121660151A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to prediction systems and methods, specifically to a precise unit prediction system and method for power load at pumped storage power station construction sites. Background Technology
[0002] In the construction industry, the forecasting and management of electricity load for pumped storage power stations has always been a critical issue. The construction of these power stations is characterized by a wide work area, diverse types of electrical equipment, large load fluctuations, and frequent changes in construction techniques. Existing electricity load forecasting methods mainly rely on single time series data or simple statistical models, making it difficult to fully utilize multi-source data (such as environmental parameters, equipment topology, and construction progress). Therefore, they suffer from deficiencies in forecast accuracy, timeliness, and adaptability. Currently, electricity load management at construction sites relies heavily on manual, experience-based scheduling. This not only leads to low accuracy in actual load forecasting but also results in delayed safety warnings and a lack of intelligent scheduling decisions. Furthermore, existing systems are significantly inadequate in handling carbon emission assessments and energy-saving optimization for construction electricity, failing to meet the requirements of green construction in the current context.
[0003] With the development of smart construction site and digital twin technologies, accurate load forecasting using multi-source sensing data and deep learning methods has become possible. However, existing deep learning models still face challenges in handling spatiotemporal multi-scale features, dynamic topological relationships, and noise interference, and current technologies and methods are insufficient to fully address these challenges. Traditional forecasting methods, due to their singularity and fixed structure, are inadequate when facing complex construction environments. While deep learning models can handle high-dimensional data, practical applications still require solutions to better integrate spatiotemporal information, handle complex dynamic relationships, and effectively reduce noise. Summary of the Invention
[0004] Purpose of the invention: The purpose of this invention is to provide a precise unit-level load prediction system for pumped storage power station construction sites to improve the accuracy, safety, adaptability, and timeliness of unit-level load prediction. On the other hand, it provides a precise unit-level load prediction method for pumped storage power station construction sites.
[0005] Technical solution: The precise unit prediction system for power load at pumped storage power station construction sites described in this invention includes: The multi-source data acquisition module is used to collect real-time power consumption, environmental parameters, equipment topology relationships, and power consumption change characteristic variables associated with the construction schedule of the target construction unit. The spatiotemporal feature matrix construction module is used to divide the pumped storage power station construction site into grid cells, and construct a feature vector for each grid cell based on data from the multi-source data acquisition module to form a spatiotemporal feature matrix. The load forecasting module is used to input the spatiotemporal feature matrix into a spatiotemporally coordinated multi-scale residual cyclic transformer model for load forecasting. The spatiotemporal feature matrix input to the spatiotemporal co-operated multi-scale residual cyclic transformer model includes: Trainable wavelet basis functions are used to convolve the input data, and spatial and temporal paths are processed in parallel to generate joint spatiotemporal features. The joint spatiotemporal features are then input into a multi-layer residual gated recurrent unit (GRU) network containing cross-layer connections for time series prediction, and the unit-level load curve for the specified future time period is output. The early warning and scheduling decision module is used to generate a list of equipment for staggered scheduling and trigger a graded early warning when the predicted value of the unit-level load curve exceeds a preset threshold.
[0006] Preferably, it also includes a multi-source sensing module, smart meters, temperature and humidity sensors, and BIM progress tracking terminals deployed at the pumped storage power station construction site.
[0007] Preferably, it also includes an edge computing node module for performing abnormal data cleaning and non-uniform sampling compensation.
[0008] Preferably, the tiered early warning system includes immediate early warning and periodic early warning, wherein: When the predicted value for a single period exceeds the limit, the immediate warning is triggered, non-critical equipment is shut down, and equipment that can be delayed is marked. When the predicted limits are exceeded for three consecutive cycles, the cycle warning is triggered, the diesel backup unit is automatically started, and suggestions for adjusting the construction plan are pushed.
[0009] The method for accurate unit prediction of power load at pumped storage power station construction sites according to the present invention includes the following steps: Collect real-time power consumption data, environmental parameters, equipment topology relationships, and power consumption change characteristics associated with the construction schedule of the target construction unit; The pumped storage power station construction site is divided into multiple grid units, and a feature vector is constructed for each grid unit to form a spatiotemporal feature matrix. The spatiotemporal feature matrix is input into the spatiotemporal co-operated multi-scale residual cyclic transformer model to predict the unit-level load curve for a specified future period. The spatiotemporal co-operated multi-scale residual cyclic transformer model is executed as follows: The input data is convolved using trainable wavelet basis functions, and the steady-state load component and transient noise component are separated by a gated recursive unit. Parallel processing of spatial and temporal paths generates joint spatiotemporal features; The joint spatiotemporal features are input into a multi-layer residual gated cyclic unit (GRU) network containing cross-layer connections for time series prediction, and the unit-level load curve for the specified future time period is output. When the predicted value of the unit-level load curve exceeds the preset threshold, a list of equipment off-peak scheduling is generated and a graded early warning is triggered.
[0010] Preferably, the feature vector of each grid cell includes: active power, reactive power, temperature, humidity, equipment type code, and construction stage code.
[0011] Preferably, the spatial path calculates spatial influence weights based on the device topology using a graph attention network (GAT); the temporal path extracts temporal dependency features using a ProbSparse multi-head attention mechanism, and the output features of the two paths are weighted and fused using dynamic fusion parameters to generate joint spatiotemporal features.
[0012] Preferably, the multi-layer residual gated cyclic unit network structure includes at least 8 layers of GRUs, and a cross-layer connection is set between the 1st layer and the 5th layer to directly transmit the device basic power consumption characteristics output by the 1st layer to the input of the 5th layer; The spatiotemporal co-operated multi-scale residual cyclic transformer model is trained and optimized using a joint loss function that combines quantile loss function and temporal smoothing constraint.
[0013] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. By introducing a spatiotemporal collaborative multi-scale residual cyclic transformer model, the temporal dynamic characteristics and spatial distribution features of power load can be effectively integrated, significantly improving the accuracy of unit-level load prediction; 2. Through a noise reduction mechanism combining trainable wavelet basis functions and gated recursive units, steady-state load and transient noise can be accurately separated in complex construction environments, thereby improving data quality; 3. By dynamically updating the equipment topology relationship through cable length and impedance ratio, the impact of changes in the location of construction equipment on load distribution can be reflected in real time. Attached Figure Description
[0014] Figure 1 This is a schematic diagram of the system structure of the present invention;
[0015] Figure 2 This is a schematic diagram of the principle framework of the present invention;
[0016] Figure 3 This is a schematic diagram of the STC-MRRT model structure of the present invention;
[0017] Figure 4 This is a schematic diagram of the system principle of the feature processing unit of the present invention. Detailed Implementation
[0018] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.
[0019] As attached Figure 1 As shown in the figure, this embodiment provides a precise unit prediction system for power load at a pumped storage power station construction site, including a multi-source data acquisition module, a spatiotemporal feature matrix construction module, a load prediction module, and an early warning and dispatch decision module.
[0020] For details, see attached. Figure 2 As shown: The multi-source data acquisition module is used to collect real-time power consumption, environmental parameters, equipment topology relationships, and power consumption change characteristics associated with the construction schedule of the target construction unit.
[0021] Among them, real-time power consumption data includes active power and reactive power, environmental parameters include temperature and humidity, equipment topology relationships, and power consumption change characteristic variables associated with the construction schedule can be understood as variables containing process power consumption characteristic parameters.
[0022] The spatiotemporal feature matrix construction module is used to divide the pumped storage power station construction site into grid cells. Based on the data from the multi-source data acquisition module, a feature vector is constructed for each grid cell to form a spatiotemporal feature matrix.
[0023] Specifically, the feature vector constructed for each grid cell can be represented as: [active power, reactive power, temperature, humidity, equipment type code, construction stage code].
[0024] The load forecasting module is used to input the spatiotemporal feature matrix into the spatiotemporal co-operated multi-scale residual cyclic transformer model for load forecasting. The spatiotemporal co-operated multi-scale residual cyclic transformer model is STC-MRRT.
[0025] The load forecasting module includes: The multi-scale recursive noise reduction unit uses trainable wavelet basis functions to convolve the input data and separates steady-state load from transient noise through a gated recursive unit. The spatiotemporal collaborative attention computation unit processes spatial and temporal paths in parallel and outputs joint features through a dynamic fusion parameter α. Specifically, the spatial path calculates spatial influence weights based on device topology using a graph attention network (GAT). The temporal path extracts temporal dependency features through a ProbSparse multi-head attention mechanism and uses a dynamic fusion parameter to weight and fuse the output features of the two paths to output joint spatiotemporal features.
[0026] The residual gated recurrent prediction unit (GRU) inputs joint spatiotemporal features into a multi-layer residual gated recurrent unit (GRU) network containing cross-layer connections for time series prediction, and outputs the unit-level load curve for a specified future period. Specifically, the multi-layer residual gated recurrent unit network structure contains at least 8 GRU layers, and cross-layer connections are set between the 1st and 5th layers, directly transmitting the device basic power consumption features output from the 1st layer to the input of the 5th layer. The spatiotemporal co-operated multi-scale residual recurrent transformer model is trained and optimized using a joint loss function that combines quantile loss function and time series smoothing constraint.
[0027] For example, the unit-level load curves for the next 15 minutes to 24 hours are output.
[0028] The early warning and scheduling decision module generates a list of equipment for staggered peak scheduling and triggers a tiered early warning when the predicted value of the unit-level load curve exceeds the preset threshold of the transformer capacity; for example, the early warning and scheduling decision module is triggered when the predicted value exceeds 85% of the transformer capacity.
[0029] Optionally, the system also includes a multi-source sensing module, smart meters, temperature and humidity sensors, and BIM progress tracking terminals deployed at the pumped storage power station construction site.
[0030] Optionally, it also includes an edge computing node module for performing abnormal data cleaning and non-uniform sampling compensation; specifically, the edge computing node module integrates an NVIDIA Jetson AGX Orin chip, and the compensation formula for non-uniform sampling compensation is as follows: ; Where P(t, x) represents the compensated power value at time t and position x; t represents the current time point; x represents the current position coordinates; Δt represents the time interval, i.e., the difference between the previous sampling time and the current time; P(t-Δt, x) represents the historical power value at time t-Δt and position x. Represents the coordinates of adjacent positions of the current position x; Represents time t and adjacent positions The power value at that location.
[0031] The edge computing node module is also configured with an anomaly cleaning strategy: (1) Set the fluctuation threshold coefficient to 1.8 for the excavation stage and set the coefficient to 1.2 for the lining stage; (2) When the data exceeds the threshold, it is marked as a transient event or a real anomaly in the device start-up and shutdown log.
[0032] Optionally, the tiered early warning system includes immediate early warning and periodic early warning, wherein: When a single-cycle forecast value exceeds the limit, an immediate warning is triggered, non-critical equipment is shut down, and equipment that can be delayed is marked. When the prediction exceeds the limit for three consecutive cycles, a cycle warning is triggered, the diesel standby unit is automatically started, and suggestions for adjusting the construction plan are pushed.
[0033] This invention also provides a method for accurate unit prediction of power load at pumped storage power station construction sites, comprising the following steps: S1. Collect real-time power consumption data, environmental parameters, equipment topology relationships, and power consumption change characteristics associated with the construction schedule of the target construction unit.
[0034] The methods for constructing device topology relationships include: (1) Generate an adjacency matrix based on cable length and impedance ratio The expression is as follows: .
[0035] in, Indicates the cable lengths of devices i and j. This represents the impedance ratio.
[0036] (2) Dynamically update the topology weights. Recalculate the adjacency matrix based on the changes in the location of the construction equipment at each set unit time interval.
[0037] S2. Divide the pumped storage power station construction site into multiple grid units, and construct a feature vector for each grid unit to form a spatiotemporal feature matrix.
[0038] The feature vector of each grid cell includes: active power, reactive power, temperature, humidity, equipment type code, and construction stage code.
[0039] S3. Input the spatiotemporal feature matrix into the spatiotemporal co-operated multi-scale residual cyclic transformer model to predict the unit-level load curve for a specified future period. Among them, the spatiotemporal co-operational multi-scale residual cyclic transformer model is executed as follows: S31. The input data is convolved using trainable wavelet basis functions, and the steady-state load component and transient noise component are separated by a gated recursive unit.
[0040] Specifically, adaptive convolution is performed on the input data using Daubechies6 or Coiflet3 wavelet basis functions, with the mathematical form as follows:
[0041] Where x(t) represents the input time series data, and α represents the scaling parameter (also called the scaling factor). denoted by , represents the complex conjugate of the wavelet basis function, used to perform inner product operation with the input data x(t) to realize the multi-scale decomposition of the signal, and b represents the translation parameter.
[0042] An exponential attenuation factor λ=0.85 is applied to the high-frequency noise components, while the low-frequency trend components are enhanced.
[0043] S32 processes spatial and temporal paths in parallel and outputs joint features through dynamic fusion parameter α.
[0044] Specifically, the dynamic fusion parameter α is generated through a learnable weight matrix, satisfying the following formula: .
[0045] Wherein, α represents the dynamic fusion parameter, which is used to control the fusion weight of spatial path features and temporal path features in the final joint features; α represents the activation function, which maps the input value to the range of 0 to 1, so that the value of alpha can reasonably represent the fusion ratio of the two features; W represents the learnable weight matrix, which is continuously optimized during the model training process and is used to perform weighted transformation on the concatenated vector of spatial and temporal features; Spatial_feat represents the feature vector extracted from the spatial path, reflecting the relevant information of the data in the spatial dimension; Temporal_feat represents the feature vector extracted from the temporal path, reflecting the temporal characteristics of the data changing over time; || represents the vector concatenation operation, which concatenates the spatial feature vector and the temporal feature vector into a new vector for subsequent weighting and other operations; b represents the bias term, which is also a learnable parameter used to adjust the weighted result, making the model more flexible and better at fitting.
[0046] When calculating the influence weights of the spatial path graph attention network (GAT), a construction stage coefficient β is introduced to correct the attention score: .
[0047] Where β represents the construction stage coefficient, which is used to correct the attention score when the spatial path graph attention network (GAT) calculates the influence weight of the equipment; stage_code represents the construction stage code, with different code values corresponding to different construction stages, thereby reflecting the differences in the influence weight of the construction stage on the equipment.
[0048] S33. Input the joint spatiotemporal features into a multi-layer residual gated cyclic unit (GRU) network containing cross-layer connections for time series prediction, and output the unit-level load curve for the specified future time period.
[0049] Specifically, cross-links are added between GRU layers, directly connecting the device's basic power consumption characteristics output from layer 1 to the input of layer 5; quantile loss function and temporal smoothing constraints are used for joint optimization. .
[0050] Where L represents the joint loss function, used to evaluate the deviation between the model's predictions and the actual values, guiding model optimization; T represents the total length of the predicted time series, i.e., how many time periods' load values to predict; t represents the time index, from 1 to T, representing different time periods; q represents the quantile parameter, used to control the weight of the quantile loss, affecting the model's attention to the prediction errors of different quantiles; This represents the actual load value at time t. This represents the predicted load value at time t. This represents the predicted load value at time (t - 1). This represents the L2 norm, which is used here to calculate the difference between predicted values at adjacent times, thereby implementing time-series smoothing constraints and making the predicted load curve more stable.
[0051] S4. When the predicted value of the unit-level load curve exceeds the preset threshold, a list of equipment off-peak scheduling is generated and a graded early warning is triggered.
[0052] Optionally, an online learning mechanism is also included, triggering incremental training every 24 hours. If the prediction bias increases by 15% consecutively, model reconstruction is initiated, using Bayesian optimization to adjust hyperparameters in the following space: ; ; .
Claims
1. A precise unit prediction system for power load at a pumped storage power station construction site, characterized in that, include: The multi-source data acquisition module is used to collect real-time power consumption, environmental parameters, equipment topology relationships, and power consumption change characteristic variables related to the construction schedule of the target construction unit. The spatiotemporal feature matrix construction module is used to divide the pumped storage power station construction site into grid cells, and construct a feature vector for each grid cell based on data from the multi-source data acquisition module to form a spatiotemporal feature matrix. The load forecasting module is used to input the spatiotemporal feature matrix into a spatiotemporally coordinated multi-scale residual cyclic transformer model for load forecasting. The spatiotemporal feature matrix input to the spatiotemporal co-operated multi-scale residual cyclic transformer model includes: Trainable wavelet basis functions are used to convolve the input data, and spatial and temporal paths are processed in parallel to generate joint spatiotemporal features. The joint spatiotemporal features are then input into a multi-layer residual gated recurrent unit (GRU) network containing cross-layer connections for time series prediction, and the unit-level load curve for the specified future time period is output. The early warning and scheduling decision module is used to generate a list of equipment for staggered scheduling and trigger a graded early warning when the predicted value of the unit-level load curve exceeds a preset threshold.
2. The system according to claim 1, characterized in that, It also includes a multi-source sensing module, smart meters, temperature and humidity sensors, and BIM progress tracking terminals deployed at the pumped storage power station construction site.
3. The system according to claim 1, characterized in that, It also includes an edge computing node module, which is used to perform abnormal data cleaning and non-uniform sampling compensation.
4. The system according to claim 1, characterized in that, The tiered early warning system includes immediate early warning and periodic early warning, wherein: When the predicted value for a single period exceeds the limit, the immediate warning is triggered, non-critical equipment is shut down, and equipment that can be delayed is marked. When the predicted limits are exceeded for three consecutive cycles, the cycle warning is triggered, the diesel backup unit is automatically started, and suggestions for adjusting the construction plan are pushed.
5. A method for accurate unit prediction of power load at a pumped storage power station construction site, characterized in that, Includes the following steps: Collect real-time power consumption data, environmental parameters, equipment topology relationships, and power consumption change characteristics associated with the construction schedule of the target construction unit; The pumped storage power station construction site is divided into multiple grid units, and a feature vector is constructed for each grid unit to form a spatiotemporal feature matrix. The spatiotemporal feature matrix is input into the spatiotemporal co-operated multi-scale residual cyclic transformer model to predict the unit-level load curve for a specified future period. The spatiotemporal co-operated multi-scale residual cyclic transformer model is executed as follows: The input data is convolved using trainable wavelet basis functions, and the steady-state load component and transient noise component are separated by a gated recursive unit. Parallel processing of spatial and temporal paths generates joint spatiotemporal features; The joint spatiotemporal features are input into a multi-layer residual gated cyclic unit (GRU) network containing cross-layer connections for time series prediction, and the unit-level load curve for the specified future time period is output. When the predicted value of the unit-level load curve exceeds the preset threshold, a list of equipment off-peak scheduling is generated and a graded early warning is triggered.
6. The method according to claim 5, characterized in that, The feature vector of each grid cell includes: active power, reactive power, temperature, humidity, equipment type code, and construction stage code.
7. The method according to claim 5, characterized in that, The spatial path calculates spatial influence weights based on the device topology using a graph attention network (GAT); the temporal path extracts temporal dependency features using a ProbSparse multi-head attention mechanism, and the output features of the two paths are weighted and fused using dynamic fusion parameters to generate joint spatiotemporal features.
8. The method according to claim 5, characterized in that, The multi-layer residual gated cyclic unit network structure includes at least 8 layers of GRUs, and a cross-layer connection is set between the 1st and 5th layers to directly transmit the device basic power consumption characteristics output from the 1st layer to the input of the 5th layer. The spatiotemporal co-operated multi-scale residual cyclic transformer model is trained and optimized using a joint loss function that combines quantile loss function and temporal smoothing constraint.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 5 to 8.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 5 to 8.