Power operation control method and system based on big data pre-compression grid processing

Abstract

The present invention discloses a power operation control method and system based on big data pre-compression and grid processing. The method includes the following steps: S1, pre-compression through data block mechanism and dynamic dictionary compression; S2, grid processing through three-dimensional grid division and anomaly detection; S3, optimal state mining of the results of grid processing based on genetic algorithm and particle swarm optimization algorithm; S4, based on the results of electricity price prediction and the results of optimal state mining, determining the activation rate and corresponding output power parameters of each device; S5, performing visual output. The system is used to support the operation of this method. The present invention uses distributed hash tables and double-layered dictionary coding to quickly compress massive data and reduce computing load; perform multi-dimensional grid division and hybrid algorithm calculation to accurately locate and obtain optimal operating parameters; perform electricity price prediction and generate decision plans and visualize them, making the control process more intuitive.

Description

Technical Field

[0001] The present invention relates to the field of digital management and control technology in the energy industry, and in particular to a power operation management and control method and system based on big data pre-compression grid processing. Background Art

[0002] Energy management and control involves monitoring and optimizing the production, distribution, and use of energy through a range of technical means and management strategies to improve energy efficiency, reduce energy consumption, lower energy costs, and minimize environmental impact. The management and control process involves applying various technologies and management methods to monitor and optimize energy generation, distribution, and consumption, thereby improving energy efficiency, reducing unnecessary energy use, and mitigating negative environmental impacts.

[0003] Against the backdrop of the continuous development of the electricity trading market, new adjustments and changes have occurred in the production and operation models of the power industry. Currently, power generation companies are facing many urgent problems. For example, profit margin compression: Intensified competition in electricity market transactions has further squeezed the profits of power generation companies. Companies urgently need to reduce costs and improve competitiveness by improving operational efficiency and other means; unscientific operation methods: Current operation methods mostly rely on human control and lack scientific and effective means, making it difficult to achieve precise regulation and optimization of power operation; historical optimal state is difficult to perceive: Due to the lack of effective technical means, power generation companies cannot accurately perceive the optimal state in historical operating conditions, making it difficult to fully utilize historical experience to improve current operating levels; data processing difficulties: During the operation of power generation companies, massive amounts of historical data are generated, but existing technical means cannot effectively process and utilize this data, resulting in a waste of data resources.

[0004] In response to existing problems in the electric power industry, the document with publication number CN110991722A discloses a method for predicting electric power load. This method predicts electric power load based on a hierarchical parallel Bayesian neural network, using a large amount of historical data to establish a Bayesian neural network and a model for predicting actual grid load. However, this prediction method is difficult to adapt to the complex periodicity of the electric power load itself and lacks an intuitive display of the results. In addition, the prediction content of the load data is relatively simple, resulting in the prediction results not accurately reflecting future electric power load data, and the predicted results are difficult to directly apply to actual electric power load regulation. In addition, if the amount of historical data is too large, it will also lead to an excessively high computational load. Summary of the Invention

[0005] In response to the above problems, the purpose of the present invention is to propose a power operation control method and system based on big data pre-compression grid processing. Through distributed hash tables and double-layered dictionary coding, it can achieve rapid compression of massive data and significantly reduce the computing load; it also performs multi-dimensional grid division and hybrid algorithm calculation to accurately locate data and obtain optimal operating parameters, thereby improving the accuracy of prediction; it also completes electricity price prediction and thermal power unit operation plan based on LSTM neural network and decision tree, and then performs visual display to make the control process more intuitive.

[0006] This is achieved through the following technical solutions:

[0007] First, a power operation control method based on big data pre-compression grid processing is proposed, which includes the following steps: S1. Pre-compressing the to-be-processed data of the thermal power unit through a data block mechanism and dynamic dictionary compression; S2. Gridding the pre-compressed data through three-dimensional grid division and anomaly detection; S3. Optimal state mining of the results of the gridding processing based on genetic algorithm and particle swarm optimization algorithm, including: using the results of the gridding processing as a data grid model, randomly combining the operating parameters in each grid and using them as each initial population individual of the genetic algorithm, and then simulating the natural evolution process to screen out each target operating parameter combination; using the position of the initial population individual corresponding to each target operating parameter combination as the initial position, obtaining the global optimal solution through the particle swarm optimization algorithm, and taking the global optimal solution as the result of optimal state mining; S4. Using LSTM neural network to predict electricity prices; based on the results of electricity price prediction and the results of optimal state mining, using decision tree to determine the activation rate of each device in the thermal power unit and the corresponding output power parameters; S5. Visualizing the electricity price prediction, output power parameters and equipment activation rate based on natural language processing and generative adversarial network.

[0008] Preferably, during pre-compression in step S1, the data to be processed is first divided into multiple data blocks, each of which is quickly indexed and located using a distributed hash table. A dynamic dictionary with a dual-layer architecture is then constructed for each data block for compression, to remove initial noise, which includes repeated characters or repeated numerical sequences. The dual-layer architecture includes a base data layer and an index mapping layer. The base data layer is used to map data with initial noise into short codes and also to map data without initial noise into long codes. The index mapping layer is used to index and locate each long code in the base data layer. Through dual-layer compression using data segmentation and a dynamic dictionary, data size can be effectively compressed while taking data location into account.

[0009] Preferably, when dividing the data to be processed into multiple data blocks, the data to be processed is divided according to time series, spatial dimension, and data volume. When dividing according to time series, if the data to be processed is real-time operating data of a thermal power unit, each data block is divided according to minutes, and the granularity rule for the division is set to D1 minutes, where D1 ≥ 15. If the data to be processed is historical operating data of a thermal power unit, each data block is divided according to months or quarters, and the granularity rule is set in the same unit. When dividing according to spatial dimension, data generated by the same device or system is divided into one data block. When dividing according to data volume, a data volume threshold D2 is first set, and multiple data blocks are divided based on the data volume threshold D2. Multi-dimensional block division according to time, space, and data volume is consistent with the actual data characteristics of thermal power units.

[0010] Preferably, during gridding in step S2, the pre-compressed data is first partitioned into multiple dimensions based on time, operating parameter, and device status. A first boundary is set in the time dimension based on multiple fixed time periods, a second boundary is set in the operating parameter dimension based on the rated parameters or safety thresholds of each device, and a third boundary is set in the device status dimension based on the operating mode of each device. This three-dimensional boundary partitioning provides a structured framework for the data grid model, facilitating subsequent anomaly detection and efficient execution of various algorithms.

[0011] Preferably, after completing the multidimensional boundary division, the data features within each grid are extracted to form corresponding feature vectors. Each feature vector is then fed into the Isolation Forest algorithm, which constructs multiple isolation trees by randomly selecting features and segmentation points. Based on each isolation tree, each abnormal grid and each normal grid are determined. Each abnormal grid is removed, and the set of normal grids is used as the result of the gridding process. Using the Isolation Forest algorithm facilitates rapid identification of anomalies.

[0012] Preferably, the genetic algorithm in step S3 is provided with an adaptive function ,in, 、 and are all weight adjustment coefficients that change dynamically with the evolutionary generation t, Efficiency is efficiency, Cost is cost, and Emission is emission; 、 and They are the preset inertia weights of efficiency, cost and emission respectively; when obtaining each initial population individual of the genetic algorithm, the number of initial population individuals is first adjusted using the adaptive function; then the probability of individual selection is set for each individual of the adjusted initial population , , is the average fitness of the population, is the adjustment factor, i and j are both positive integers, n corresponds to the number of individuals in the initial population, and the probability of individuals being selected is used Multiple individuals are selected from each adaptively adjusted initial population to participate in a crossover and mutation operation. Multiple crossover and mutation strategies are then applied to each selected individual to obtain each target operating parameter combination. By combining the adaptive function, individual selection probability, and crossover and mutation operations, potential conflicts between individuals can be balanced, thereby improving computational accuracy.

[0013] Preferably, the particle swarm optimization algorithm in step S3 is provided with an adaptive degree update formula ,

[0014] in, , , , is the inertia weight, is the learning factor, the subscript max represents the maximum value, the subscript min represents the minimum value, the subscript id represents the particle i in dimension d, t represents the tth iteration, and T is the maximum number of iterations. is the fitness variance, is the particle distribution density, is the maximum density, is the adjustment coefficient, r is a random number, is the individual historical optimal, is the global optimal of the group, is a preset constant, is the position of particle i in the current dimension d. When using the particle swarm optimization algorithm to obtain the global optimal solution, the initial population of individuals corresponding to each target operating parameter combination in step S3 is first divided into multiple subgroups. Each subgroup is updated using the adaptive speed update formula, and the global optimal solution is obtained based on the updated results. The structure of multiple subgroups and adaptive parameter updates can enhance global search capabilities and avoid falling into local optimal solutions.

[0015] Preferably, in step S4, the LSTM neural network adopts a multi-input parallel architecture. Each architecture separately introduces historical electricity price data, historical fuel prices, historical temperature and humidity parameters, and historical electricity supply and demand as input sequences. Each input sequence is subjected to feature conversion through an independent embedding layer. The results of each feature conversion are fused in the hidden layer, and the fused results are output through the output layer of the LSTM neural network to obtain the electricity price forecast result. Using a multi-input parallel architecture and introducing multiple input sequences can improve prediction accuracy.

[0016] Preferably, in step S5, the natural language processing adopts a hierarchical attention enhancement architecture and is provided with a knowledge fusion module. The hierarchical attention enhancement architecture includes a bottom module for semantic feature extraction and a middle module using a multi-dimensional attention mechanism. The middle module uses a multi-dimensional focusing formula to output a text description; the multi-dimensional focusing formula is , where Q, K, and V are query vector, key vector, and value vector, respectively. K is the dimension of the key vector, E is the length of the time dimension, The weight matrix for each dimension is composed of Multi, Dim, and Weights, which correspond to the time dimension, equipment dimension, and economic dimension, respectively. The knowledge fusion module constructs a knowledge graph containing multiple nodes based on the GNN neural network, and uses each node to construct the underlying module. The content corresponding to each node is equipment knowledge or electricity trading knowledge. The generative adversarial network adopts a three-layer structure of bottom layer, middle layer, and top layer. The bottom layer is used to receive the result data of the decision tree and the text description data output by natural language processing for feature mapping. The middle layer uses the feature alignment module to align and match any data in the bottom layer to image features. After the alignment and matching are completed, the top layer generates each image for display. Through the collaborative output of natural language processing and generative adversarial networks, images that are easy to display can be obtained, allowing observers to understand the situation more intuitively and accurately.

[0017] In addition, a system is proposed, which includes a data acquisition module, a preprocessing module, a pre-compression module, a grid processing module, an optimal state analysis module, an electricity price prediction module, an operation strategy generation module and a visualization display module; wherein the data acquisition module adopts a dual-channel redundant backup mechanism to transmit the to-be-processed data of the thermal power unit through two identical channels, which are used for normal operation and backup respectively; the preprocessing module is used to perform preliminary processing on the to-be-processed data transmitted by the data acquisition module, including data cleaning and format conversion, and then transmit it to the pre-compression module; the pre-compression module pre-compresses the data transmitted by the pre-processing module based on a pre-compression mechanism combining distributed hash tables and dictionary coding; the grid processing module grids the pre-compressed data through three-dimensional grid division and anomaly detection; the optimal state analysis module calculates and analyzes the results of the grid processing through genetic algorithms and particle swarm optimization algorithms to obtain the global optimal solution; the electricity price prediction module performs electricity price prediction based on the LSTM neural network; the operation strategy generation module generates the operation strategy using a decision tree based on the electricity price prediction results and the global optimal solution; the visualization display module visualizes the operation strategy based on natural language processing and generative adversarial networks.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] The technical solution of the present invention uses distributed hash tables and double-layered dictionary coding to achieve rapid compression of massive data, significantly reducing the computing load. It also performs multi-dimensional grid division and hybrid algorithm calculations to accurately locate data and obtain optimal operating parameters, thereby improving prediction accuracy. It also uses LSTM neural networks and decision trees to complete electricity price forecasts and thermal power unit operation plans, and then performs visual display to make the management and control process more intuitive. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of a power operation control method based on big data pre-compression grid processing;

[0021] Figure 2 This is a framework diagram of a power operation control system based on big data pre-compression grid processing. DETAILED DESCRIPTION

[0022] The following will be combined with the present invention Figure 1 and Figure 2 , the technical solutions in the embodiments of the present invention are described in detail.

[0023] like Figure 1 As shown in the figure, it is a flow chart of a power operation control method based on big data pre-compression and grid processing. Pre-compression is used to reduce the data volume and computing load, and the data are accurately divided and feature combined during grid processing, so as to use the hybrid algorithm to mine the optimal state. Finally, the electricity price forecast is combined with the optimal state to calculate the final operation mode of the thermal power unit and visualize it, so that the prediction process can easily cope with large amounts of data, with high accuracy and greatly improved intuitiveness.

[0024] The method specifically comprises the following steps:

[0025] S1. Pre-compress the data to be processed by the thermal power unit through data block mechanism and dynamic dictionary compression.

[0026] In this embodiment, the data to be processed can be real-time operating data of a thermal power unit or historical operating data. During pre-compression, the data to be processed is first divided into multiple data blocks based on time series, spatial dimension, and data volume. Each data block is quickly indexed and located using a distributed hash table.

[0027] When the data is divided according to time series and the data to be processed is the real-time operation data of thermal power units, each data block is divided into minutes and the granularity rule for division is set to D1 minutes, D1 ≥ 15; this is because the frequency of data generation is high during real-time operation. This setting can not only ensure that the data in each data block has high relevance and timeliness, but also achieve a balance between data volume and processing complexity.

[0028] When the data to be processed is the historical operating data of thermal power units, each data block can be divided into units of months or quarters and the granularity rules can be set in the same units; this is because historical operating data often has a large time span and a relatively slow data change rate. Dividing it by month or quarter facilitates long-term trend analysis and statistical calculations.

[0029] When partitioned according to spatial dimensions, a thermal power plant comprises multiple different devices and systems, such as boilers, steam turbines, and generators. The data generated by each device and system has distinct spatial attributes. Spatial data partitioning involves grouping data generated by the same device or system into a single data block. The corresponding granularity rules can be further refined based on the functional modules of each device or system. For example, data for a boiler's combustion system and steam-water system can be partitioned into separate data blocks to better meet the management and control requirements of different devices and systems.

[0030] When partitioning by data size, we first set a data volume threshold, D2, and then divide the data into multiple data blocks based on this threshold. This size-based partitioning ensures balanced and efficient data processing. When data reaches this threshold, it is automatically partitioned into a new data block. This partitioning method can accommodate differences in data generation rates over time, avoiding situations where disparate data block sizes can lead to excessive processing time for some blocks and insufficient processing time for others.

[0031] In this embodiment, after completing data segmentation, a dual-layered dynamic dictionary is constructed for each data block for compression to remove initial noise, which includes repeated characters or repeated numerical sequences. The dual-layered architecture includes a base data layer and an index mapping layer. The base data layer is used to map data with initial noise into short codes and also to map data without initial noise into long codes. The index mapping layer is used to locate the corresponding index for each long code in the base data layer, and the data corresponding to the short code no longer needs to be used. Through dual-layered compression of data segmentation and dynamic dictionary, data size can be effectively compressed while taking into account data location.

[0032] S2. Gridding the pre-compressed data through three-dimensional grid division and anomaly detection.

[0033] In this embodiment, when performing gridding processing, the pre-compressed data is first divided into multiple dimensions according to the time dimension, the operating parameter dimension, and the device status dimension. In the time dimension, the first boundary is set according to multiple fixed time periods, for example, divided into units of hours, days, and weeks; in the operating parameter dimension, the second boundary is set according to the rated parameters or safety thresholds of each device, for example, divided according to the corresponding data such as current, voltage, and power; in the device status dimension, the third boundary is set according to the operating mode of each device, for example, the device status is divided into normal, faulty, and maintenance. The three-dimensional boundary division can provide a structured framework for the data grid model, facilitating subsequent anomaly detection and the efficient execution of various algorithms.

[0034] After completing the multidimensional boundary demarcation, data features within each grid are extracted, such as the mean and standard deviation of the time dimension, extreme values ​​of operating parameters, rates of change, and distribution ratios of equipment states, to form corresponding feature vectors. Each feature vector is then fed into the isolation forest algorithm, which constructs multiple isolated trees by randomly selecting features and split points. Each isolated tree calculates the path length of the data. The shorter the path, the more likely the data is an anomaly. Therefore, an anomaly score is calculated: an anomaly score is calculated for each grid based on the path length of the data within each isolated tree. A higher score indicates a greater likelihood of an anomaly within the grid. Next, an abnormal grid is identified: a threshold for the anomaly score is set. When a grid's score exceeds the threshold, it is identified as an anomaly. For example, the threshold can be determined by cross-validating different grids, with each grid with a score above the 95th percentile being marked as an anomaly. This allows each isolated tree to identify both abnormal and normal grids, further eliminating the impact of abnormal grids. The resulting gridding process is then used as the collection of normal grids. Using the isolation forest algorithm facilitates rapid anomaly identification.

[0035] S3. Based on the genetic algorithm and particle swarm optimization algorithm, the results of grid processing are mined for the optimal state, including: using the results of grid processing as the data grid model, combining the operating parameter combinations in each grid as each initial population individual of the genetic algorithm, and then simulating the natural evolution process to screen out each target operating parameter combination; using the position of the initial population individual corresponding to each target operating parameter combination as the initial position, obtaining the global optimal solution through the particle swarm optimization algorithm, and using the global optimal solution as the result of optimal state mining.

[0036] When using a genetic algorithm, dynamic population size adjustment is first performed: the population size is dynamically adjusted based on the number of evolutionary generations and fitness convergence during algorithm iterations. In the early stages of evolution, the population size is increased to broaden the search space. As evolution progresses, if fitness stabilizes, the population size is appropriately reduced to improve computational efficiency. Next, multi-strategy crossover and mutation are implemented: A variety of crossover and mutation strategies are employed to address the characteristics of different data grids. For example, for grids with large fluctuations in operating parameters, the mutation probability is increased and a multi-point crossover strategy is employed. For grids with relatively stable parameters, the mutation probability is reduced and a single-point crossover strategy is employed to balance local search and global exploration capabilities.

[0037] When the genetic algorithm performs iterative evolution, the adaptive function used is ,in, 、 and are weight adjustment coefficients that change dynamically with the evolutionary generation t. 、 、 Represent efficiency, cost, and emission respectively. 、 and They are the preset inertia weights of efficiency, cost and emission, which increase in the early stage of evolutionary iteration. The weight of the search is to focus on the power generation efficiency, and in the later stage of evolutionary iteration, the and weights to balance costs and emissions.

[0038] When obtaining each initial population individual of the genetic algorithm, the adaptive function is first used to adjust the number of initial population individuals; then the probability of individual selection is set for each individual in the adjusted initial population. , , is the average fitness of the population, is the adjustment factor, i and j are both positive integers, n corresponds to the number of individuals in the initial population, individuals with fitness higher than the average level have a higher probability of being selected; individuals with fitness lower than the average level have a lower probability of being selected, which can accelerate the evolution of the population. After selecting multiple individuals from each adaptively adjusted initial population, they can participate in a crossover / mutation operation. Multiple crossover / mutation strategies are used to mutate each selected individual to obtain each target operating parameter combination. By combining the adaptive function, individual selection probability, and crossover / mutation, potential conflicts between individuals can be balanced, thereby improving computational accuracy.

[0039] When performing mutation operation on any individual, the value after mutation is recorded as ,but , where if means if, otherwise means other situations opposite to if; With mean 0 and variance Normally distributed random numbers, is the scaling factor, UB is Upper Bound, UB represents X j The corresponding upper bound value, LB stands for Lower Bound, LB represents X j The corresponding lower bound value, r1 is a selected random number, and Δ represents a dynamically varying step length function, which enables the mutation operation to explore better solutions in the local neighborhood.

[0040] In this embodiment, a particle swarm optimization algorithm uses a hierarchical particle swarm structure, dividing the particle swarm into multiple subgroups, each corresponding to a different data grid or operating condition. Particles within a subgroup focus on local search, and subgroups regularly exchange information, thereby achieving efficient coordination between local and global optima and avoiding being trapped in local optima. The algorithm is equipped with an inertia weight and two learning factors, which can be dynamically adjusted based on the particle distribution density and fitness variance, thereby strengthening local search in dense particle areas and enhancing global exploration capabilities in sparse areas.

[0041] The adaptive update formula of the particle swarm optimization algorithm is: ,

[0042] in, , , , is the inertia weight, is the learning factor, the subscript max represents the maximum value, the subscript min represents the minimum value, the subscript id represents the particle i in dimension d, t represents the tth iteration, and T is the maximum number of iterations. is the fitness variance, is the particle distribution density, is the maximum density, is the adjustment coefficient, r is a random number, is the individual historical optimal, is the global optimal of the group, is a preset constant, It is also used to adjust the coefficient to control the adjustment intensity of the fitness variance to the inertia weight. is the position of particle i in the current dimension d.

[0043] When using the particle swarm optimization algorithm to obtain the global optimal solution, the initial population of individuals corresponding to each target operating parameter combination in step S3 is first divided into multiple subgroups. Each subgroup is updated using the adaptive speed update formula, and the global optimal solution is obtained based on the updated results. The structure of multiple subgroups and adaptive parameter updates can enhance global search capabilities and avoid being trapped in local optimal solutions.

[0044] S4. Use LSTM neural network to predict electricity prices. Based on the results of electricity price prediction and optimal state mining, use decision tree to determine the activation rate of each device in the thermal power unit and the corresponding output power parameters.

[0045] In this embodiment, the LSTM neural network adopts a multi-input parallel architecture. Each architecture separately introduces historical electricity price data, historical fuel prices, historical temperature and humidity parameters, and historical power supply and demand as input sequences. Each input sequence is subjected to feature transformation through a separate embedding layer. The results of each feature transformation are fused in the hidden layer and output through the output layer of the LSTM neural network to obtain the electricity price forecast. Using a multi-input parallel architecture and introducing multiple input sequences can improve prediction accuracy.

[0046] When performing fusion, the fusion function used is: ,in, 、 、 、 are the corresponding feature vectors of different input sequences after being transformed by the embedding layer, is the activation function, is the bias vector, This fusion method can make full use of the correlation between multi-dimensional data and enhance the model's ability to capture the factors affecting electricity prices.

[0047] In addition, when using LSTM neural networks, you can also incorporate an attention mechanism, which first calculates the attention weights of each time step in the input sequence , , , where U is the sequence length, h u is the hidden state at the u-th time step, is the hidden state of the last time step, exp represents the exponential operation with e as the base, is a function that computes the attention score, such as a dot product or a multilayer perceptron; the weighted hidden state is then used as the output: As a result, electricity price forecasts are closely linked to historical data, especially when processing long series of data, highlighting important information, ignoring noise, and improving forecast accuracy.

[0048] The LSTM neural network employs a multi-scale hierarchy. The bottom-level LSTM units take electricity price data at short intervals (e.g., 15 minutes) as input to capture short-term fluctuations in the electricity market. The middle-level LSTM units take hourly data as input to learn medium-term price trends. The top-level LSTM units take daily data as input to grasp long-term market price patterns. Information is transferred between layers via residual connections. The output relationship between layer l and layer l-1 is: , l is the number of layers, z is the input of layer l at the uth time step. This multi-scale hierarchical structure can analyze electricity price data from different time granularities, integrate short-term fluctuations, medium-term trends and long-term laws, and improve the model's adaptability to complex electricity price change patterns.

[0049] S5. Based on natural language processing and generative adversarial networks, visualize the electricity price forecast, output power parameters, and equipment activation rate.

[0050] In this embodiment, natural language processing adopts a hierarchical attention enhancement architecture and is provided with a knowledge fusion module. The hierarchical attention enhancement architecture includes a bottom module for semantic feature extraction and a middle module using a multi-dimensional attention mechanism. The middle module uses a multi-dimensional focusing formula to output text descriptions; the multi-dimensional focusing formula is , where Q, K, and V are query vectors, key vectors, and value vectors respectively. Multi, Dim, and Weights correspond to the time dimension, device dimension, and economic dimension respectively. K is the dimension of the key vector, E is the length of the time dimension, The softmax function is the commonly known softmax function, which is a weight matrix composed of each dimension. The knowledge fusion module constructs a knowledge graph containing multiple nodes based on the GNN neural network, using each node to construct the underlying module. The content corresponding to each node is equipment knowledge or power trading knowledge. The generative adversarial network adopts a three-layer structure of bottom layer, middle layer, and top layer. The bottom layer is used to receive the result data of the decision tree and the text description data output by natural language processing for feature mapping. The middle layer uses the feature alignment module to align and match any data in the bottom layer to image features. After the alignment and matching are completed, the top layer generates each image for display. Through the coordinated output of natural language processing and generative adversarial networks, images that are easy to display can be obtained, allowing observers to understand the situation more intuitively and accurately.

[0051] The weight matrix formed by combining each dimension is specifically quantified into three dimensions: time, equipment, and economy. The weight matrix is ​​a combined matrix obtained by weighted addition of the sub-matrices of each of the three dimensions. The dimensions of each sub-matrix remain consistent, and the weighting ratio can be freely adjusted. This weighted addition method allows the resulting results to better take into account the overall situation.

[0052] First, quantification of the temporal dimension focuses on the quantification of time scales, temporal characteristics, and temporal correlations. Its core objective is to transform temporal information into variables recognizable by the model. Quantification methods include basic quantification, derived quantification, and temporal correlation quantification. Time granularity, also known as basic quantification, is used to divide continuous time into discrete units, serving as the basic unit of quantification. For example, in the short term, granularity can be divided into minutes (e.g., 15-minute data collection) and hours (24 hours / day), with m = 1, 2, ..., 24 representing the 24 hours of a day. In the medium term, granularity can be divided into days (30 days / month) and weeks (7 days / week), with m = 1, 2, ..., 30 representing the 30 days of a month. In the long term, granularity can be divided into seasons (4 seasons / year) and years (12 months / year), with s = 1, 2, 3, 4 representing spring, summer, autumn, and winter, respectively. For example, the real-time operating data of a thermal power unit is recorded hourly, where m=1 corresponds to 0:00-1:00 and m=15 corresponds to 14:00-15:00. The power generation, coal consumption, and other data for each hour are bound to the time tag.

[0053] Temporal feature quantification, also known as derivative quantification, can extract quantitative indicators based on the "periodicity," "trend," and "peak-valley characteristics" of time. For example, time period attributes can be quantified by dividing electricity prices into "peak, valley, and flat" periods, using labels (e.g., peak time = 3, flat time = 2, and valley time = 1), or encoded using a 0-1 scale (peak time = 1, off-peak time = 0). Time intervals can be calculated by calculating the time difference between two events, e.g., "number of days between last maintenance and current operation" = current date - maintenance date (unit: days). Periodicity indicators can be quantified using "cosine / sine functions" (applicable to time series models).

[0054] Time correlation quantification is used to quantify the "dependence between current time and historical time". For example, for time series trends, the "sliding window mean" is used to represent short-term trends (such as the average load of the past 24 hours), and the "difference series" is used to represent the rate of change (such as current load minus the load of the previous hour).

[0055] Secondly, quantification of the equipment dimension focuses on quantifying the equipment's status, performance, and health. Its core objective is to reflect the actual status of the equipment through sensor data and statistical indicators. Quantification methods include real-time quantification, efficiency quantification, and long-term quantification. The quantification of equipment status parameters, also known as real-time quantification, directly quantifies the equipment's real-time operating status based on physical quantities collected by sensors. For example, core parameters such as boiler steam temperature (°C), turbine inlet steam pressure (MPa), and generator active power (MW) are directly represented by measured values ​​(e.g., 540°C, 16.7 MPa). Status labels quantify "normal / abnormal" status as a 0-1 variable (normal = 0, abnormal = 1), and fault levels are classified (minor fault = 1, major fault = 3). For example, real-time data such as the coal mill current (A) and induced draft fan speed (r / min) of a thermal power unit are directly used as input features for the equipment dimension.

[0056] Performance indicators, or efficiency metrics, are used to quantify the equipment's operating efficiency or load capacity, reflecting its output level. These indicators include load factor and power generation efficiency. Load factor is calculated as: actual power / rated power × 100% (e.g., a unit with a rated power of 600MW and a current output of 480MW has a load factor of 80%). Power generation efficiency is calculated as: power generation / fuel consumption × 100% (e.g., if 3,000 kWh of electricity is generated per ton of standard coal consumed, efficiency = 3,000 / 8,600 = 34.9%; note: the calorific value of 1 ton of standard coal is ≈ 8,600 kWh).

[0057] Health and reliability are long-term quantitative measures used to quantify equipment aging and failure risks, providing a basis for maintenance decisions. For example, failure rate is calculated as: number of failures divided by cumulative operating hours (e.g., a water pump operates for 8,000 hours per year and has three failures; failure rate = 3 / 8,000 = 0.000375 failures / hour). Remaining useful life (RUL) is estimated using degradation models (such as linear degradation or neural network prediction) to estimate the remaining normal operating time of the equipment (e.g., the remaining useful life of a steam turbine = 12,000 hours).

[0058] Furthermore, the quantification of the economic dimension focuses on the quantification of "cost," "benefit," and "economic indicators," with the core being to reflect "economic value" using monetary units or efficiency indicators. Quantification methods include cost quantification, benefit quantification, and comprehensive economic indicators. Cost quantification is a core indicator used to calculate various expenditures for equipment operation or decision-making, expressed in monetary units (such as yuan). For example, fuel cost: fuel consumption × unit price (e.g., a coal-fired unit consumes 5,000 tons of coal at a unit price of 1,000 yuan / ton, fuel cost = 5 million yuan); maintenance cost: number of maintenance times × single maintenance fee + spare parts cost (e.g., three maintenance sessions per month at 20,000 yuan each, spare parts cost 50,000 yuan, monthly maintenance cost = 3 × 2 + 5 = 110,000 yuan); start-up and shutdown costs: the converted cost of fuel consumption and equipment wear during the start-up and shutdown of a thermal power unit (e.g., a single startup cost ≈ 50,000 yuan).

[0059] Quantified benefits are used to calculate the economic returns from operations or decisions. For example, power generation benefits are calculated as: power generation × grid-connected electricity price (e.g., if 100,000 kWh of power is generated during a certain period and the price is 0.5 yuan / kWh, the benefits are 50,000 yuan). Peak shaving benefits are calculated as the additional subsidies received when participating in grid peak shaving (e.g., if deep peak shaving is performed with a subsidy of 0.2 yuan per kWh and the peak shaving capacity is 20,000 kWh, the benefits are 4,000 yuan).

[0060] Comprehensive economic indicators, which quantify efficiency, can be used to reflect overall economic performance using metrics such as "unit cost" and "return on investment." For example, unit power generation cost is calculated as: total operating cost / total power generation (e.g., if monthly total costs are 8 million yuan and power generation is 20 million kWh, unit cost = 0.4 yuan / kWh); return on investment (ROI): (annual net income / total equipment investment) × 100% (e.g., if the total unit investment is 3 billion yuan and the annual net income is 300 million yuan, the ROI = 10%).

[0061] A generative adversarial network (GAN) consists of a generator and multiple discriminators. The generator generates a visualization image through layered processing based on input conditions. The generated image is then evaluated by multiple discriminators, and the discriminators output scores that are fed back to the generator. The generator then uses the backpropagation algorithm to adjust various image-related parameters based on the feedback from each discriminator to optimize the generated image. For example, if a local discriminator indicates inaccurate labeling of a power generation curve, the generator adjusts the relevant parameters, correcting the curve and labeling information. This process repeats until the generated image meets quality requirements. Once the generated image passes the evaluation of all discriminators and meets the set quality standards, it is output to ensure clarity and intuitiveness. The generated visualizations include power generation-price relationship diagrams and dynamic flow charts of equipment operating status. These intuitively display data such as the optimal number and parameters of operating equipment, operating mode, and predicted electricity prices, providing clear and easy-to-understand decision-making guidance for plant operators. This enables the efficient conversion of thermal power unit operation strategies from data to visual, user-friendly outputs, significantly improving the understandability and practicality of the operation strategies.

[0062] like Figure 2 Figure 2 shows the framework of a power operation control system based on big data pre-compression and grid processing. The system includes a data acquisition module, a pre-processing module, a pre-compression module, a grid processing module, an optimal state analysis module, an electricity price forecasting module, an operation strategy generation module, and a visualization module. The data acquisition module utilizes a dual-channel redundant backup mechanism, transmitting the pending data from thermal power units through two identical channels, one for normal operation and the other for backup. The pre-processing module performs preliminary processing on the pending data transmitted by the data acquisition module, including data cleaning and format conversion, before transmitting it to the pre-compression module. The pre-compression module pre-compresses the data transmitted by the pre-processing module using a pre-compression mechanism that combines distributed hash tables and dictionary encoding. The grid processing module performs grid processing on the pre-compressed data through three-dimensional meshing and anomaly detection. The optimal state analysis module calculates and analyzes the grid processing results using genetic algorithms and particle swarm optimization algorithms to obtain the global optimal solution. The electricity price forecasting module uses an LSTM neural network to predict electricity prices. The operation strategy generation module uses a decision tree to generate an operation strategy based on electricity price forecasts and the global optimal solution. This strategy modifies the activation rate of each device in the thermal power unit and the corresponding output power of each enabled device. The visualization module uses natural language processing and generative adversarial networks to visualize the operation strategy.

[0063] In summary, the present invention uses distributed hash tables and double-layered dictionary coding to achieve rapid compression of massive data and significantly reduce the computational load. It also performs multi-dimensional grid division and hybrid algorithm calculations to accurately locate data and obtain optimal operating parameters, thereby improving prediction accuracy. It also uses LSTM neural networks and decision trees to complete electricity price forecasts and thermal power unit operation plans, and then performs visual display, making the management and control process more intuitive and significantly progressive.

[0064] The above embodiments are only for illustrating the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A power operation control method based on big data pre-compression grid processing, characterized in that: The steps include: S1. Pre-compress the data to be processed by the thermal power unit through data block mechanism and dynamic dictionary compression; S2. Gridding the pre-compressed data through three-dimensional gridding and anomaly detection, including: first, performing multi-dimensional boundary partitioning on the pre-compressed data according to the time dimension, the operating parameter dimension, and the device status dimension; after completing the multi-dimensional boundary partitioning, extracting the data features within each grid to form corresponding feature vectors; then inputting each feature vector into the isolation forest algorithm, and constructing multiple isolation trees by randomly selecting features and segmentation points; determining each abnormal grid and each normal grid based on each isolation tree, eliminating each abnormal grid, and using the set of each normal grid as the result of the gridding processing; S3. Based on the genetic algorithm and particle swarm optimization algorithm, the results of the grid processing are subjected to optimal state mining, including: using the results of the grid processing as the data grid model, randomly combining the operating parameters within each grid and using them as each initial population individual of the genetic algorithm, then simulating the natural evolution process to screen out each target operating parameter combination; using the position of the initial population individual corresponding to each target operating parameter combination as the initial position, obtaining the global optimal solution through the particle swarm optimization algorithm, and using the global optimal solution as the result of the optimal state mining; S4. Use LSTM neural network to predict electricity prices. Based on the results of electricity price prediction and optimal state mining, use decision tree to determine the activation rate and corresponding output power parameters of each device in the thermal power unit. S5. Based on natural language processing and generative adversarial networks, visualize the electricity price forecast, output power parameters, and equipment activation rate.

2. The power operation control method based on big data pre-compression grid processing according to claim 1 is characterized in that: During pre-compression in step S1, the data to be processed is first divided into multiple data blocks, and each data block is quickly indexed and located through a distributed hash table; then, a dynamic dictionary with a dual-layer architecture is constructed for each data block for compression to remove initial noise, which includes repeated characters or repeated numerical sequences; wherein the dual-layer architecture includes a basic data layer and an index mapping layer, the basic data layer is used to map the data of the initial noise into short codes, and is also used to map the data of non-initial noise into long codes, and the index mapping layer is used to perform corresponding index positioning for each long code in the basic data layer.

3. The power operation control method based on big data pre-compression grid processing according to claim 2 is characterized in that: When dividing the data to be processed into multiple data blocks, the data to be processed is divided according to time series, spatial dimension and data size; Among them, when dividing by time series, if the data to be processed is the real-time operation data of thermal power units, each data block is divided into minutes and the granularity rule when dividing is set to D1 minutes, D1 ≥ 15; if the data to be processed is the historical operation data of thermal power units, each data block is divided into months or quarters and the granularity rule is set in the same unit; When dividing according to the spatial dimension, the data generated by the same device or the same system is divided into a data block; When dividing according to the data volume, a data volume threshold D2 is first set, and multiple data blocks are divided based on the data volume threshold D2.

4. The power operation control method based on big data pre-compression grid processing according to claim 1 is characterized in that: When performing multi-dimensional boundary division in step S2, the first boundary is set in the time dimension according to multiple fixed time periods, the second boundary is set in the operating parameter dimension according to the rated parameters or safety thresholds of each device, and the third boundary is set in the device status dimension according to the operating mode of each device.

5. The power operation control method based on big data pre-compression grid processing according to claim 1 is characterized in that: In step S3, the genetic algorithm is set with an adaptive function ,in, 、 and are weight adjustment coefficients that change dynamically with the evolutionary generation t, Efficiency is efficiency, Cost is cost, and Emission is emission. 、 and They are the preset inertia weight of efficiency, the preset inertia weight of cost and the preset inertia weight of emission; When obtaining each initial population individual of the genetic algorithm, the adaptive function is first used to adjust the number of initial population individuals; then the probability of individual selection is set for each individual in the adjusted initial population. , , is the average fitness of the population, is the adjustment factor, i and j are both positive integers, n corresponds to the number of individuals in the initial population, and the probability of individuals being selected is used Multiple population individuals are selected from each initial population individual after adaptive adjustment to participate in the crossover mutation operation; multiple crossover mutation strategies are used to perform mutation operations on each selected population individual to obtain each target operating parameter combination.

6. The power operation control method based on big data pre-compression grid processing according to claim 5 is characterized in that: In step S3, the particle swarm optimization algorithm is set with an adaptive degree update formula. ,in, , , , is the inertia weight, is the learning factor, the subscript max represents the maximum value, the subscript min represents the minimum value, the subscript id represents the particle i in dimension d, t represents the tth iteration, and T is the maximum number of iterations. is the fitness variance, is the particle distribution density, is the maximum density, is the adjustment coefficient, r is a random number, is the individual historical optimal, is the global optimal of the group, is a preset constant, is the position of particle i in the current dimension d; When obtaining the global optimal solution through the particle swarm optimization algorithm, the initial population individuals corresponding to each target operating parameter combination in step S3 are first divided into multiple subgroups, and each subgroup is updated using the adaptive speed update formula, and the global optimal solution is obtained based on the updated results.

7. The power operation control method based on big data pre-compression grid processing according to claim 1 is characterized in that: In step S4, the LSTM neural network adopts a multi-input parallel architecture. Each architecture introduces historical electricity price data, historical fuel prices, historical temperature and humidity parameters, and historical electricity supply and demand as input sequences. Each input sequence is subjected to feature conversion through an independent embedding layer. The results of each feature conversion are fused in the hidden layer, and the fused results are output through the output layer of the LSTM neural network to obtain the result of electricity price prediction.

8. The power operation control method based on big data pre-compression grid processing according to claim 1 is characterized in that: In step S5, the natural language processing adopts a hierarchical attention enhancement architecture and is provided with a knowledge fusion module. The hierarchical attention enhancement architecture includes a bottom module for semantic feature extraction and a middle module using a multi-dimensional attention mechanism. The middle module uses a multi-dimensional focusing formula to output text descriptions; the multi-dimensional focusing formula is , where Q, K, and V are query vector, key vector, and value vector, respectively. K is the dimension of the key vector, E is the length of the time dimension, The weight matrix for each dimension is composed of Multi, Dim, and Weights, which correspond to the time dimension, equipment dimension, and economic dimension respectively. The knowledge fusion module builds a knowledge graph containing multiple nodes based on the GNN neural network, and uses each node to build the underlying module. The content corresponding to each node is equipment knowledge or power transaction knowledge. The generative adversarial network adopts a three-layer structure of bottom layer, middle layer and top layer. The bottom layer is used to receive the result data of the decision tree and the text description data output by natural language processing for feature mapping; the middle layer aligns and matches any data in the bottom layer into image features through the feature alignment module; after completing the alignment and matching, the top layer generates each image for display.

9. A power operation control system based on big data pre-compression and grid processing, which is operated by a power operation control method based on big data pre-compression and grid processing according to any one of claims 1 to 8, characterized in that: The system includes a data acquisition module, a pre-processing module, a pre-compression module, a grid processing module, an optimal state analysis module, an electricity price prediction module, an operation strategy generation module and a visualization display module; The data acquisition module adopts a dual-channel redundant backup mechanism to transmit the data to be processed by the thermal power unit through two identical channels, one for normal operation and the other for backup. The pre-processing module is used to perform preliminary processing on the data to be processed transmitted by the data acquisition module, including data cleaning and format conversion, and then transmit it to the pre-compression module; The pre-compression module pre-compresses the data transmitted by the pre-processing module based on a pre-compression mechanism that combines distributed hash tables and dictionary coding; The grid processing module performs grid processing on the pre-compressed data through three-dimensional grid division and anomaly detection; The optimal state analysis module calculates and analyzes the results of grid processing through genetic algorithms and particle swarm optimization algorithms to obtain the global optimal solution; The electricity price prediction module predicts electricity prices based on the LSTM neural network; The operation strategy generation module uses a decision tree to generate an operation strategy based on the results of electricity price forecasting and the global optimal solution; The visualization display module is based on natural language processing and generative adversarial networks to visualize the operation strategy.

Images