Transformer intelligent power saving method and system based on multi-source data fusion

Abstract

The invention provides a transformer intelligent power saving method and system based on multi-source data fusion, and belongs to the technical field of power systems. According to the method, multi-source heterogeneous data such as meteorological conditions, industrial production and resident habits are integrated to construct a space-time correlation feature set. A correlation coefficient of hour-level and day-level load characteristics is quantized by means of a long-short-term memory network, a high-correlation scene is screened through a preset correlation coefficient threshold, coupling strength between historical loads and influence factors is quantized in combination with a Pearson correlation coefficient, and key influence factors are extracted to eliminate interference. A power load prediction model is constructed based on a random forest algorithm to output a predicted value, the predicted value is converted into an instant demand fluctuation vector, finally, transformer operation optimization parameters are dynamically generated through a fuzzy control algorithm, accurate matching of the operation state and the power demand is achieved, the power supply quality and reliability are guaranteed, meanwhile, the power grid energy saving efficiency is remarkably improved, and the power supply cost is reduced. And a scientific scheme is provided for intelligent power saving of the transformer.

Description

Technical Field

[0001] This application relates to the field of power system technology, specifically to a transformer intelligent energy-saving method and system based on multi-source data fusion. Background Technology

[0002] Currently, the power system, as a crucial infrastructure of the national economy, directly impacts energy security and economic development through its operational efficiency. Transformers, as core equipment in power transmission, bear the critical responsibility of voltage transformation and power distribution; optimizing their operation is of decisive significance for energy conservation and emission reduction across the entire power grid. Current transformer operation and dispatching primarily rely on historical experience and fixed patterns for management, a traditional approach that reveals significant shortcomings when facing complex and ever-changing power demand environments. Existing dispatching schemes often rely on static configurations based on preset parameters, lacking the ability to respond sensitively to real-time changes in power demand. This results in transformers maintaining high power consumption even during low-load periods, leading to substantial energy waste.

[0003] The information disclosed in the background section is only for enhancing the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] In view of this, this application provides a transformer intelligent power saving method and system based on multi-source data fusion, which can reduce the power consumption of transformers.

[0005] In a first aspect, embodiments of this application provide a transformer intelligent power-saving method based on multi-source data fusion. The method includes: acquiring multi-source heterogeneous data, including meteorological condition data, industrial production data, and residential habit data; determining a spatiotemporal correlation feature set based on the multi-source heterogeneous data; determining a correlation coefficient between hourly fluctuation feature vectors and daily periodic change feature vectors based on a long short-term memory network and the spatiotemporal correlation feature set; determining whether the correlation coefficient exceeds a preset correlation coefficient threshold; if it exceeds the preset correlation coefficient threshold, acquiring historical power load data and power load influencing factor data, and applying the Pearson correlation coefficient... The algorithm determines the coupling strength between historical power load data and power load influencing factor data; it determines whether the coupling strength exceeds a preset coupling strength threshold. If it exceeds the preset coupling strength threshold, the current power load influencing factor data is taken as a key influencing factor; based on the random forest algorithm, a power load demand prediction model is constructed according to the key power load influencing factors; the multi-source heterogeneous data is input into the power load demand prediction model to output a predicted power load demand value; an instantaneous demand fluctuation vector is determined based on the predicted power load demand value; and based on the fuzzy control algorithm, transformer operating state optimization parameters are determined according to the instantaneous demand fluctuation vector.

[0006] Secondly, embodiments of this application provide a transformer intelligent power-saving system based on multi-source data fusion. This system includes: an acquisition module, a first determination module, a second determination module, a first judgment module, a second judgment module, a construction module, a third determination module, a fourth determination module, and a fifth determination module. The acquisition module is used to acquire multi-source heterogeneous data, including meteorological condition data, industrial production data, and residential habit data. The first determination module is used to determine a spatiotemporal correlation feature set based on the multi-source heterogeneous data. The second determination module is used to determine the correlation coefficient between hourly fluctuation feature vectors and daily periodic change feature vectors based on the spatiotemporal correlation feature set using a long short-term memory network. The first judgment module is used to determine whether the correlation coefficient exceeds a preset correlation coefficient threshold. If the threshold is exceeded, historical power load data and power load influencing factor data are acquired, and the correlation coefficient between historical power load data and power load influencing factors is determined based on the Pearson correlation coefficient algorithm. The system comprises five modules: a first module for determining the coupling strength between factor data; a second module for determining whether the coupling strength exceeds a preset coupling strength threshold, and if so, for which the current power load influencing factor data is taken as a key influencing factor; a third module for constructing a power load demand forecasting model based on the random forest algorithm and the key influencing factors; a fourth module for inputting multi-source heterogeneous data into the power load demand forecasting model to output the power load demand forecast value; a fifth module for determining the instantaneous demand fluctuation vector based on the power load demand forecast value; and a sixth module for determining the transformer operating state optimization parameters based on the instantaneous demand fluctuation vector using a fuzzy control algorithm.

[0007] This application provides a transformer intelligent power-saving method and system based on multi-source data fusion. By comprehensively integrating heterogeneous data from multiple sources such as meteorological conditions, industrial production, and residential habits, it overcomes the limitations of traditional scheduling that relies on only a single data source. Through spatiotemporal correlation analysis of multi-dimensional data, a complete correlation feature system covering time, space, and characteristic dimensions is constructed, effectively solving the problem of insufficient multi-source data fusion in traditional methods. By deeply mining the potential correlation between hourly instantaneous fluctuations and daily periodic changes using Long Short-Term Memory networks and quantifying the correlation coefficient, hierarchical correlation patterns across time scales are accurately identified, solving the technical pain point of traditional analysis's difficulty in capturing the differentiated characteristics of multi-factor coupling effects. Based on the Pearson correlation coefficient algorithm, the coupling strength between historical power load and various influencing factors is quantified, key influencing factors are screened, and a load demand prediction model is constructed using the random forest algorithm, significantly improving the relevance and accuracy of the prediction model and avoiding interference from irrelevant factors in the prediction results. By inputting real-time multi-source heterogeneous data into the prediction model to output load demand forecasts, and further transforming these forecasts into an instantaneous demand fluctuation vector reflecting the trend and intensity of load increases and decreases, the system achieves sensitive capture of real-time electricity demand changes, breaking free from the constraints of traditional static configuration. Finally, a fuzzy control algorithm is used to transform the fluctuation vector into transformer operating state optimization parameters, achieving dynamic and precise matching between transformer operating state and actual electricity demand. This effectively solves the problem of energy waste caused by high-power operation during low-load periods. While ensuring power supply quality and reliability, it significantly improves the energy-saving efficiency of the power grid, providing efficient and reliable technical support for intelligent power saving in transformers. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present application of a transformer intelligent power-saving method based on multi-source data fusion. Figure 2 This is a flowchart illustrating a transformer intelligent energy-saving method based on multi-source data fusion, provided in another exemplary embodiment of this application. Figure 3 This is a flowchart illustrating a transformer intelligent power-saving method based on multi-source data fusion, provided in another exemplary embodiment of this application. Figure 4 This is a flowchart illustrating a transformer intelligent power-saving method based on multi-source data fusion, provided in another exemplary embodiment of this application. Figure 5 This is a flowchart illustrating a transformer intelligent power-saving method based on multi-source data fusion, provided in another exemplary embodiment of this application. Figure 6 This is a flowchart illustrating a transformer intelligent power-saving method based on multi-source data fusion, provided in another exemplary embodiment of this application. Figure 7 This is a flowchart illustrating a transformer intelligent power-saving method based on multi-source data fusion, provided in another exemplary embodiment of this application. Figure 8 This is a flowchart illustrating a transformer intelligent power-saving method based on multi-source data fusion, provided in another exemplary embodiment of this application. Detailed Implementation

[0010] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this application will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this application.

[0011] The terms “a,” “one,” and “the” are used to indicate the existence of one or more elements / components / etc.; the terms “including” and “having” are used to indicate an open-ended inclusion and that other elements / components / etc. may exist in addition to those listed. The terms “first” and “second” are used only as markers and are not a limitation on the number of objects.

[0012] Currently, the power system, as a crucial infrastructure of the national economy, directly impacts energy security and economic development through its operational efficiency. Transformers, as core equipment in power transmission, bear the critical responsibility of voltage transformation and power distribution; optimizing their operation is of decisive significance for energy conservation and emission reduction across the entire power grid. Current transformer operation and dispatching primarily rely on historical experience and fixed patterns for management, a traditional approach that reveals significant shortcomings when facing complex and ever-changing power demand environments. Existing dispatching schemes often rely on static configurations based on preset parameters, lacking the ability to respond sensitively to real-time changes in power demand. This results in transformers maintaining high power consumption even during low-load periods, leading to substantial energy waste.

[0013] For example, transformer load demand is influenced by multiple factors, including weather conditions, industrial production cycles, and residential electricity consumption habits. These factors exhibit complex coupling relationships across different time scales. Summer heatwaves trigger surges in air conditioning loads, while adjustments to industrial production plans cause load fluctuations at specific times. The combined demand pattern is difficult to accurately grasp through single-dimensional analysis. Further complicating matters, this multi-factor coupling effect exhibits differentiated characteristics at different time granularities. A hierarchical correlation mechanism exists between hourly instantaneous fluctuations and daily / weekly periodic changes, making it difficult for traditional analytical methods to effectively identify and quantify this cross-timescale correlation.

[0014] Therefore, how to construct a load demand forecasting mechanism that can accurately capture the spatiotemporal correlation characteristics between multi-source data and achieve precise matching between transformer operating status and actual power demand has become a technical problem that needs to be solved in intelligent power saving technology for transformers.

[0015] This application provides a transformer intelligent power-saving method and system based on multi-source data fusion, such as... Figure 1 The illustrated method is a smart energy-saving method for transformers based on multi-source data fusion. This method may include the following steps: Step S110: Acquire multi-source heterogeneous data, which includes meteorological condition data, industrial production data, and resident habit data; Step S120: Determine the spatiotemporal correlation feature set based on multi-source heterogeneous data; Step S130: Based on the long short-term memory network, determine the correlation coefficient between the hourly fluctuation feature vector and the daily periodic change feature vector according to the spatiotemporal correlation feature set; Step S140: Determine whether the correlation coefficient exceeds the preset correlation coefficient threshold. If it exceeds the preset correlation coefficient threshold, obtain historical power load data and power load influencing factor data, and determine the coupling strength between historical power load data and power load influencing factor data based on the Pearson correlation coefficient algorithm. Step S150: Determine whether the coupling strength exceeds the preset coupling strength threshold. If it is determined that the coupling strength exceeds the preset coupling strength threshold, then the current power load influencing factor data will be used as the key influencing factor. Step S160: Based on the random forest algorithm, construct an electricity load demand prediction model according to the key influencing factors of electricity load; Step S170: Input multi-source heterogeneous data into the power load demand forecasting model to output the power load demand forecast value; Step S180: Determine the instantaneous demand fluctuation vector based on the electricity load demand forecast; Step S190: Based on the fuzzy control algorithm, determine the transformer operating state optimization parameters according to the real-time demand fluctuation vector.

[0016] According to the intelligent power-saving method for transformers based on multi-source data fusion provided in this application, the method can acquire multi-source heterogeneous data, including meteorological data, industrial production data, and residential habit data; determine a spatiotemporal correlation feature set based on the multi-source heterogeneous data; determine the correlation coefficient between hourly fluctuation feature vectors and daily periodic change feature vectors based on the spatiotemporal correlation feature set using a long short-term memory network; determine whether the correlation coefficient exceeds a preset correlation coefficient threshold; if it exceeds the preset correlation coefficient threshold, acquire historical power load data and power load influencing factor data, and determine the coupling strength between historical power load data and power load influencing factor data based on the Pearson correlation coefficient algorithm; determine whether the coupling strength exceeds a preset coupling strength threshold; if it exceeds the preset coupling strength threshold, take the current power load influencing factor data as the key influencing factor; construct a power load demand prediction model based on the key influencing factors of power load using a random forest algorithm; input the multi-source heterogeneous data into the power load demand prediction model to output the power load demand prediction value; determine the instantaneous demand fluctuation vector based on the power load demand prediction value; and determine the transformer operating state optimization parameters based on the instantaneous demand fluctuation vector using a fuzzy control algorithm.

[0017] This method, by deeply integrating multi-source heterogeneous data such as meteorological conditions, industrial production, and residential habits, breaks through the limitations of traditional transformer dispatching that relies solely on single data or historical experience, laying a comprehensive data foundation for accurate load demand analysis. Through spatiotemporal correlation mining of multi-source heterogeneous data, a spatiotemporal correlation feature set is constructed, clearly revealing the dynamic correlation patterns among meteorological, industrial, and residential data, solving the problems of insufficient multi-source data fusion and difficulty in capturing complex correlations in traditional methods. By leveraging a Long Short-Term Memory (LSTM) network for deep processing of the spatiotemporal correlation feature set, the correlation coefficient between hourly fluctuation feature vectors and daily periodic change feature vectors is accurately quantified, successfully identifying the hierarchical correlation mechanism between hourly instantaneous fluctuations and daily periodic changes, overcoming the technical difficulty of traditional analysis methods failing to effectively analyze the differentiated characteristics of cross-timescale coupling effects. Next, a preset correlation coefficient threshold is used for screening, further analyzing only scenarios corresponding to high correlation features, and combining the Pearson correlation coefficient algorithm to quantify the coupling strength between historical power load data and various power load influencing factor data. Finally, a preset coupling strength threshold is used to screen key influencing factors, eliminating interference from irrelevant or weakly correlated factors. Based on the random forest algorithm, a power load demand forecasting model is constructed with key influencing factors as the core. Leveraging the ensemble learning advantage of the algorithm, the model fully explores the interactive effects of various key influencing factors on load demand, significantly improving the accuracy of load demand forecasting and avoiding the large prediction deviations of traditional single-model approaches. Simultaneously, real-time collected multi-source heterogeneous data is input into the power load demand forecasting model, outputting accurate power load demand forecasts. These forecasts are further transformed into real-time demand fluctuation vectors containing key information such as load increase / decrease trends and fluctuation intensity, enabling sensitive capture of real-time electricity demand changes and overcoming the limitations of traditional dispatching based on static configuration of preset parameters and lag in response. Finally, through a fuzzy control algorithm, combined with the real-time demand fluctuation vector, transformer operating state optimization parameters are dynamically generated. This allows for flexible adjustment of transformer operating configurations (such as voltage regulation and load allocation) based on real-time load changes, achieving precise matching between transformer operating states and actual electricity demand. This effectively solves the problem of energy waste caused by transformers maintaining high power consumption during low-load periods in traditional dispatching. While ensuring the power supply quality and operational reliability of the power grid, this significantly improves the energy-saving efficiency of the entire power grid, providing a scientific and efficient technical solution for intelligent power saving in transformers.

[0018] The steps of the intelligent power-saving method for transformers based on multi-source data fusion provided in this application are described in detail below: In one embodiment of this application, step S110 involves acquiring multi-source heterogeneous data, including meteorological condition data, industrial production data, and residential habit data. Specifically, meteorological condition data is a key external factor affecting transformer load demand, including temperature, relative humidity, wind speed, precipitation, sunshine duration, and air pressure. Automated data collection is achieved through meteorological monitoring stations deployed in different geographical areas and IoT meteorological sensors. Some data can be obtained by connecting to the national meteorological data platform or the open interface of local meteorological bureaus to acquire historical and real-time data. The collection frequency is set according to load analysis needs, typically collecting real-time data once per hour, generating 24 time-series records daily to ensure coverage of the daily variation patterns of parameters such as temperature and humidity. For example, taking the meteorological condition data collection around an industrial park in summer as an example, the real-time meteorological condition data at 12:00 on July 15, 2024 are: temperature 35.2℃, relative humidity 62%, wind speed 2.5 m / s, precipitation 0 mm, cumulative sunshine duration 5.8 hours, and air pressure 1008 hPa.

[0019] Industrial production data is a core driver of transformer load demand, including real-time production capacity, production load, equipment operating status (e.g., operation / shutdown / maintenance), energy consumption monitoring values ​​(e.g., hourly power consumption), and emission concentrations (e.g., particulate matter concentration, exhaust gas emissions) of industrial enterprises. Industrial IoT sensors can be deployed on production equipment and energy consumption meters to collect equipment operating parameters and energy consumption data in real time. Industrial data acquisition focuses on key nodes throughout the entire industrial production process. It supports integration with industrial control systems (such as SCADA systems) to export related data such as production plans and process adjustments. For example, taking the industrial production data collection of a chemical enterprise as an example, the real-time data from 14:00 to 14:01 on July 15, 2024 is as follows: production capacity 85%, production load 1250 kilowatts (kW), equipment operation status "Reactor No. 1 in operation, Reactor No. 2 on standby", energy consumption monitoring value 1248 kW, emission concentration 25 mg / m³. This set of data can reflect the dynamic impact of industrial production activities on transformer load.

[0020] Residential habit data is key data reflecting the electricity demand of residents, including electricity consumption patterns (such as electricity consumption every 15 minutes recorded by smart meters), travel trajectories (such as transportation card swipe records and location changes in mobile phone signaling data), water consumption data (such as average daily water consumption recorded by smart water meters), and consumption behavior (such as the distribution of shopping times in community supermarkets). Among these, residential electricity consumption pattern data is the core item directly related to transformer load. For example, taking the collection of residential habit data in a certain city as an example, the residential electricity consumption pattern data on July 15, 2024, is as follows: average electricity consumption of 120 kWh from 7:00 to 7:15 and average electricity consumption of 850 kWh from 19:00 to 19:15. Travel trajectory data shows that 75% of the residents in this residential area are out during the 7:00-8:00 time period on weekdays, and 68% are returning home during the 18:00-19:00 time period. This set of data can be used to reflect the correlation between residents' travel-return patterns and the residential electricity load of transformers.

[0021] In one embodiment of this application, step S120, determining the spatiotemporal correlation feature set based on multi-source heterogeneous data, further includes the following steps: Figure 2 As shown, the specific content is as follows: Step S210: Based on the K-means clustering algorithm, cluster the multi-source heterogeneous data to determine the spatial partitioning characteristics of different spatial regions; Step S220: Based on the sliding time window algorithm, extract time sequence features from the spatial partition features to determine the time window features of different time windows; Step S230: Based on the Pearson correlation coefficient algorithm, determine the cross-latitude correlation coefficient matrix according to the time window characteristics; Step S240: Determine the spatiotemporal correlation feature set based on the cross-latitudinal correlation coefficient matrix.

[0022] Specifically, the K-means clustering algorithm divides multi-source heterogeneous data into K non-overlapping spatial regions by calculating the similarity between them (based on the similarity of geographical coordinates and data features, such as the mean of meteorological parameters, the mean of industrial load, and the mean of residential electricity consumption). Data within each region exhibits similar load impact patterns (e.g., industrial park regions, residential areas, and commercial areas), thus enabling region-based load characteristic analysis and avoiding misjudgments caused by cross-regional data contamination. For industrial park regions, clustering analysis requires combining regional industrial load time-series data acquired through industrial data collection to ensure the clustering results accurately match the spatial distribution and load characteristics of industrial production. Furthermore, a sliding time window algorithm is used to set a fixed-length time window and step size, segmenting and dynamically updating the multi-source heterogeneous data for each spatial region. This extracts time-series statistical features (such as mean, variance, and rate of change) and abnormal fluctuation indicators within the window, thereby capturing the periodic patterns (such as daily and weekly) and instantaneous fluctuations (such as sudden increases in industrial load) of the data in the time dimension, solving the problems of fragmented and irregular time-series data. The Pearson correlation coefficient algorithm is used to quantify the strength of linear associations between data from different dimensions (values ​​range from -1 to 1; the closer the absolute value is to 1, the stronger the correlation; positive correlation means that when one increases, the other also increases, and negative correlation means the opposite). Here, "cross-dimensional" refers to pairwise combinations of three dimensions: meteorological conditions data, industrial production data, and residents' habits data. By calculating the Pearson correlation coefficient of core indicators, a cross-dimensional correlation coefficient matrix is ​​constructed to clearly present the association patterns between multi-source data, solving the problem of the inability to quantify dimensional associations. The spatiotemporal association feature set is a three-dimensional integration of spatial partition features, time window features, and cross-dimensional correlation coefficients. Each feature corresponds to a specific spatial region + a specific time window + multi-dimensional association patterns, forming structured input data that can be directly used for subsequent modeling, ensuring no omissions in space, time, or features.

[0023] For example, taking the processing of multi-source heterogeneous data from a small to medium-sized city as an example, the K-means clustering algorithm (K=3) is used to divide the multi-source data from 20 monitoring points in the city into three spatial areas according to geographical location and load patterns: the East City Industrial Park, the West City Residential Area, and the City Center Commercial Area. The East City Industrial Park had an average industrial load of 1000kW and residential electricity consumption of 280kW over the past 7 days, while the West City Residential Area had an average residential electricity consumption of 750kW and an industrial load of only 40kW over the past 7 days. Then, a sliding time window algorithm (7-day window length, 1-day step) is used to extract temporal features. For example, within the window of July 1st-7th in the East City Industrial Park, industrial data collection revealed an average industrial load of 1020kW, which increased to 1180kW on July 3rd due to temporary production increases. Finally, the Pearson correlation coefficient algorithm is used to calculate cross-dimensional correlations. The correlation coefficient between temperature and residential electricity consumption in the West City Residential Area is 0.83 (higher air conditioning electricity consumption during high temperatures), and the correlation coefficient between temperature and industrial load is 0.21 (lower industrial activity has less impact), forming a correlation coefficient matrix. Finally, by integrating this information, features such as spatial region: Chengxi residential area; time window: July 1-7; average temperature 24℃, average residential electricity consumption 750kW, and temperature-residential electricity consumption Pearson correlation coefficient 0.83 are generated. By summarizing all combinations, a complete spatiotemporal correlation feature set is obtained.

[0024] In the above method, the K-means clustering algorithm is used to divide multi-source heterogeneous data into regions with similar load patterns according to spatial dimensions, which solves the problem of feature ambiguity caused by cross-regional data mixing. Secondly, the sliding time window algorithm sets window parameters to effectively capture the periodic patterns and abnormal fluctuations of time series data, avoiding misjudgment of patterns caused by time series fragmentation. Finally, the Pearson correlation coefficient algorithm quantifies the cross-dimensional correlation strength between meteorological condition data, industrial production data, and resident habit data, and clearly presents the correlation pattern through the cross-dimensional correlation coefficient matrix, which solves the pain point of traditional methods being unable to quantify the correlation of multi-dimensional data.

[0025] In one embodiment of this application, step S130, which determines the correlation coefficient between the hourly fluctuation feature vector and the daily periodic change feature vector based on the spatiotemporal correlation feature set using a long short-term memory network, further includes the following steps: Figure 3 As shown, the specific content is as follows: Step S310: Based on the sliding time window algorithm, obtain the hourly fluctuation feature vector and the daily periodic change feature vector according to the spatiotemporal correlation feature set; Step S320: Determine the multi-dimensional spatiotemporal feature matrix based on the hourly fluctuation feature vector and the daily periodic change feature vector; Step S330: Based on the long short-term memory network, determine the coupling mode between the hourly fluctuation feature vector and the daily periodic change feature vector according to the multi-dimensional spatiotemporal feature matrix; Step S340: Based on the Pearson correlation coefficient algorithm, determine the correlation coefficient between the hourly fluctuation feature vector and the daily periodic change feature vector according to the coupling mode.

[0026] Specifically, the sliding time window algorithm can extract hourly fluctuation feature vectors (reflecting instantaneous load changes over a short period) and daily periodic change feature vectors (reflecting the periodic patterns of load throughout the day) from a spatiotemporal correlation feature set by setting window lengths and step sizes with differentiated time granularities. This ensures accurate matching of analysis needs at different time scales. For example, the sliding time window length can be set to 1 hour and the step size to 15 minutes. The load deviation rate within the window (the difference between the current 15-minute load and the hourly average / hourly average × 100%) can be extracted as the core feature. This can generate 96 data points per day (24 hours × 4 15-minute intervals), forming an hourly fluctuation feature vector with a dimension of 96. For industrial park scenarios, this feature vector needs to incorporate hourly production load fluctuation data from industrial data collection to accurately reflect instantaneous load changes caused by equipment start-up and shutdown and capacity adjustments. The sliding time window length can be set to 1 day, and the step size to 1 day (the window is updated once a day). The morning peak load value (average value from 6:00 to 8:00), evening peak load value (average value from 19:00 to 21:00), and valley load value (average value from 0:00 to 5:00) within the window are extracted as core features. 3×7=21 data points can be generated in 7 days, forming a daily periodic change feature vector with dimension 21.

[0027] The multidimensional spatiotemporal feature matrix is ​​a structured data formed by aligning and integrating hourly fluctuation feature vectors and daily periodic change feature vectors over time. The row dimension corresponds to the daily time unit (7 days), and the column dimension corresponds to the hourly feature + daily feature, ensuring that the long short-term memory network can learn the temporal relationship between the two types of vectors simultaneously. For example, with 1 day as the row unit, each row includes the hourly fluctuation feature vector (96 elements) of the day + the daily periodic feature of the day (3 elements: morning peak, evening peak, and valley), ultimately forming a multidimensional spatiotemporal feature matrix with the number of rows = the number of days in the time window (7) and the number of columns = 96 + 3 = 99.

[0028] Long Short-Term Memory (LSTM) networks construct a two-layer LSTM structure (corresponding to hourly and daily features respectively) to learn the correlation patterns between two types of vectors in a multi-dimensional spatiotemporal feature matrix. Ultimately, this identifies the distribution pattern (i.e., coupling pattern) of hourly fluctuations within a daily cycle, solving the problem that traditional methods cannot resolve cross-timescale correlations. For example, the upper LSTM (128 neurons) processes hourly fluctuation features (columns 1-96), while the lower LSTM (64 neurons) processes daily cycle features (columns 97-99). An attention mechanism assigns 1.2 times the weight to features during peak hourly periods (e.g., 19:00-21:00), strengthening the correlation learning during key time periods.

[0029] For example, taking the spatiotemporal correlation feature set (time range of 3 days) of a residential community as an example, the sliding time window algorithm is used to extract hourly fluctuation feature vectors with a 1-hour window and a 15-minute step size. For example, 96 data points are collected each day, and the deviation rate of the load from the hourly average is recorded for each 15-minute period (e.g., deviation of +3% at 8 am and +10% at 7 pm). Daily cycle change feature vectors are extracted with a 1-day window and a 1-day step size, recording the morning peak (8 am) of 700kW, the evening peak (7 pm) of 900kW, and the valley value (2 pm) of 200kW. Then, the two types of vectors are integrated into a multi-dimensional spatiotemporal feature matrix with 3 rows (3 days) and 99 columns (96 hourly deviation rates + 3 daily peak and valley values), with each row corresponding to the complete features of 1 day. Then, the matrix data is learned using a long short-term memory network, and it is found that 80% of the fluctuation peaks with hourly deviation rates exceeding +8% are concentrated in the evening peak period of 7 pm to 9 pm in the daily cycle, forming a coupling pattern of hourly peaks focusing on daily evening peaks. Finally, the correlation between hourly evening peak fluctuations and daily evening peak loads was calculated using the Pearson correlation coefficient algorithm, yielding a correlation coefficient of 0.86, proving that there is a strong positive correlation between the two.

[0030] In the above method, hourly and daily feature vectors are extracted through differentiated window parameters to ensure that the two types of vectors accurately reflect the load change patterns at different time scales, avoiding the omission of instantaneous fluctuations or ambiguity of periodic features caused by traditional fixed sampling. Secondly, the two types of vectors are integrated into a multi-dimensional spatiotemporal feature matrix, realizing the structuring and time alignment of cross-scale data, providing standardized and coherent input data for the Long Short-Term Memory (LSTM) network, and solving the problem of difficult modeling of mixed multi-scale data. Thirdly, the temporal correlation learning capability of the LSTM network is utilized to accurately identify the coupling pattern between hourly fluctuations and daily cycles (such as peak concentration periods), solving the technical pain point that traditional analysis methods cannot analyze hierarchical correlations across time scales. Finally, the correlation coefficient in the coupling pattern is quantified using the Pearson correlation coefficient algorithm, providing a quantifiable basis for subsequent judgment of whether the correlation coefficient exceeds the preset correlation coefficient threshold, avoiding the bias of subjective experience judgment. This ensures the accuracy and reliability of the correlation between hourly and daily load characteristics, providing crucial cross-time scale correlation support for the accuracy of transformer load demand forecasting and subsequent operational status optimization.

[0031] In one embodiment of this application, in step S140, it is determined whether the correlation coefficient exceeds a preset correlation coefficient threshold. If it exceeds the preset correlation coefficient threshold, historical power load data and power load influencing factor data are acquired, and the coupling strength between the historical power load data and the power load influencing factor data is determined based on the Pearson correlation coefficient algorithm. Specifically, the preset correlation coefficient threshold is an effective correlation judgment standard set based on industry standards and historical experience for power load analysis. The Pearson correlation coefficient ranges from -1 to 1, where an absolute value ≥ 0.7 is defined as a strong correlation. Therefore, this step sets the preset correlation coefficient threshold to 0.7. If the correlation coefficient between the hourly fluctuation feature vector and the daily periodic change feature vector exceeds 0.7, it indicates that there is a strong cross-time scale correlation between the two types of vectors. Under effective correlation, it is necessary to further clarify which factors have a more significant impact on historical power load. Historical power load data (reflecting historical patterns of load changes) and data on factors influencing power load (including core dimensions such as meteorology, residential, and industrial factors) can be obtained. The industrial dimension data needs to be based on historical production load and capacity change records accumulated from industrial data collection. Then, the linear correlation strength (i.e., coupling strength) between the two can be calculated based on the Pearson correlation coefficient algorithm to quantify the degree of influence of each factor on the load.

[0032] For example, taking a residential community scenario as an example, the correlation coefficient between the hourly fluctuation feature vector and the daily periodic change feature vector is 0.86 > 0.7, which indicates that the scenario is a valid scenario with significant cross-time scale correlation. Next, the hourly historical power load data of the residential community over the past year was obtained (e.g., load value of 910kW at 19:00 on July 1, 2023, and load value of 210kW at 2:00 on December 1, 2023), as well as the historical data of three types of power load influencing factors during the same period (meteorological conditions influencing factor is hourly temperature data, e.g., temperature of 32℃ at 19:00 on July 1, 2023; residents' habits influencing factor is hourly electricity usage duration data, e.g., 45 minutes / household during the same period; industrial activities influencing factor is hourly load data of surrounding small factories, e.g., 120kW during the same period). Then, based on the Pearson correlation coefficient algorithm, the coupling strength between the historical power load data and the three types of influencing factor data was calculated. Finally, the coupling strength between temperature and historical power load was 0.81, the coupling strength between residents' electricity usage duration and historical power load was 0.75, and the coupling strength between factory load and historical power load was 0.32.

[0033] In one embodiment of this application, in step S150, it is determined whether the coupling strength exceeds a preset coupling strength threshold. If it is determined that the coupling strength exceeds the preset coupling strength threshold, the current power load influencing factor data is taken as a key influencing factor. Specifically, the preset coupling strength threshold is 0.7. The coupling strength of meteorological conditions (temperature) is 0.81, the coupling strength of residential habits (electricity consumption duration) is 0.75, and the coupling strength of industrial activities (factory load) is 0.32. The three types of coupling strengths are compared with the preset coupling strength threshold of 0.7. Among them, the coupling strength of temperature (0.81) > 0.7 and the coupling strength of residential electricity consumption duration (0.75) > 0.7, both of which meet the condition of exceeding the preset coupling strength threshold. Therefore, the corresponding power load influencing factor data (historical temperature data and historical residential electricity consumption duration data) are determined as key influencing factors. The coupling strength of factory load (0.32) < 0.7 does not meet the threshold requirement and is determined to be a non-key influencing factor.

[0034] In one embodiment of this application, steps S160-S170 involve constructing an electricity load demand forecasting model based on key influencing factors of electricity load using a random forest algorithm; inputting multi-source heterogeneous data into the electricity load demand forecasting model to output predicted electricity load demand values; and further including the following steps: Figure 4 As shown, the specific content is as follows: Step S410: Based on the random forest algorithm, train the key influencing factors of power load to determine the importance scores of the key influencing factors of power load; Step S420: Determine whether the importance score exceeds the preset score threshold; Step S430: If the score exceeds the preset threshold, the key influencing factors of the current power load will be used as prediction variables. Step S440: Determine the set of predictor variables based on the predictor variables; Step S450: Construct an electricity load demand forecasting model based on the set of forecast variables.

[0035] Specifically, the random forest algorithm constructs multiple independent decision trees, statistically analyzes the node splitting process of each tree, and ultimately uses the proportion of the total reduction in node impurity caused by this factor across all decision trees as the importance score of the key influencing factor on electricity load. A higher score indicates a greater contribution of this factor to electricity load demand forecasting and a stronger explanatory power for load changes. Using the set of predictor variables as input and historical electricity load data as output, the random forest model is retrained, and the model parameters are optimized through cross-validation to finally construct an electricity load demand forecasting model.

[0036] For example, a random forest algorithm was used to train the system on key factors influencing electricity load in residential communities (temperature, a meteorological factor, and electricity usage duration, a resident habit factor). The system consisted of 500 decision trees, a maximum depth of 15 layers per tree, and a minimum number of split samples of 10. Training data was extracted from hourly training data (7200 records, each containing temperature, electricity usage duration, and corresponding historical electricity load values) accumulated over the past 10 months. The final statistical results showed that temperature had an importance score of 0.42 and electricity usage duration had an importance score of 0.35. These scores were compared to a preset threshold of 0.3, and both were found to exceed this threshold. Therefore, temperature and electricity usage duration were selected as predictor variables. By integrating these two predictor variables, a set of predictor variables was formed that included temperature (hourly values) and electricity usage duration (hourly values), and was time-aligned with historical electricity load data. Using the set of predictor variables as input and historical electricity load data as output, the model was retrained with the same random forest parameters. Parameters were optimized using 5-fold cross-validation. After training, the loss function (mean squared error) converged to 0.0025. Validation using the remaining two months of test data (1440 records) showed a mean absolute error (MAE) of 3.2%. For example, if the input is July 1, 2024, at 19:00: temperature 32℃, electricity usage time 45 minutes / household, the model outputs a predicted electricity load demand of 920kW, along with a 95% confidence interval [890kW, 950kW]. This completes the construction of the electricity load demand prediction model for this residential community.

[0037] In the above method, the importance scores of key influencing factors of power load are quantified based on the random forest algorithm, transforming the subjective judgment of variable importance in traditional methods into data-driven objective quantitative results, avoiding the omission of core variables or the introduction of redundant variables due to experience bias. Secondly, predictive variables are screened by setting a preset score threshold, further refining the core predictive variables from the key influencing factors, which significantly reduces the computational complexity of model training, while avoiding noise introduced by low-contribution variables, laying the foundation for model accuracy. Finally, the random forest power load demand prediction model built based on the set of predictive variables, with the advantage of ensemble learning (multiple decision trees voting to reduce the risk of overfitting), achieves high-precision prediction on the test set and can output confidence intervals to reflect the reliability of prediction, solving the pain points of weak generalization ability and low accuracy of traditional single models (such as linear regression), ensuring that the prediction link of the entire transformer intelligent power saving method is scientific and reliable.

[0038] In one embodiment of this application, step S180, determining the instantaneous demand fluctuation vector based on the electricity load demand forecast, further includes the following steps: Figure 5 As shown, the specific content is as follows: Step S510: Determine the characteristics of meteorological fluctuations, industrial load fluctuations, and residential electricity consumption fluctuations based on the electricity load demand forecast. Step S520: Obtain the weights corresponding to meteorological fluctuation characteristics, industrial load fluctuation characteristics, and residential electricity consumption fluctuation characteristics, respectively; Step S530: Based on the weights corresponding to meteorological fluctuation characteristics, industrial load fluctuation characteristics, and residential electricity consumption fluctuation characteristics, perform weighted calculations on the meteorological fluctuation characteristics, industrial load fluctuation characteristics, and residential electricity consumption fluctuation characteristics, and use the weighted calculation results as the real-time demand fluctuation vector.

[0039] Specifically, the predicted electricity load demand is the result of the combined effects of three factors: meteorological conditions, industrial load, and residential electricity consumption. It is necessary to break down the fluctuation characteristics of each factor. For example, in terms of meteorological fluctuation characteristics: historical data shows that when the temperature remains at 25℃ (neutral temperature, with no additional impact on residential electricity consumption), the average load during this period (19:00-20:00) is 850kW; meteorological fluctuation characteristic = predicted electricity load demand - baseline load excluding temperature influence = 920kW - 850kW = 70kW (positive fluctuation, indicating that the current temperature of 32℃ is higher than the neutral temperature, driving up air conditioning electricity consumption, thus increasing the load). In terms of industrial load fluctuation characteristics: historical data shows that when surrounding small factories are shut down (no industrial load), the average load during this period is 910kW; industrial load fluctuation characteristic = 920kW - 910kW = 10kW (positive fluctuation, indicating that factories are producing only a small amount, having a slight impact on the load). Regarding the fluctuation characteristics of residential electricity consumption: Historical data shows that when residents have no additional electricity consumption (only basic lighting and socket power), the average load during this period is 800kW. The residential electricity consumption fluctuation characteristic = 920kW - 800kW = 120kW (positive fluctuation, indicating that residents turning on air conditioners and kitchen appliances after returning home is the core factor driving the load increase). The weights for meteorological fluctuation characteristics are set at 0.3, industrial load fluctuation characteristics at 0.2, and residential electricity consumption fluctuation characteristics at 0.5. The instantaneous demand fluctuation vector is calculated as follows: 70kW × 0.3 + 10kW × 0.2 + 120kW × 0.5 = 83kW. Since all three types of fluctuations are positive, the final instantaneous demand fluctuation vector is determined as [Amplitude: 83kW, Direction: Positive fluctuation, Meteorological contribution 25.3%, Industrial contribution 2.4%, Residential contribution 72.3%]. This presents the overall intensity, trend, and contribution of each factor in load fluctuations. This provides a structured and interpretable input basis for determining the transformer operating status optimization parameters (e.g., when the residential contribution is high, voltage parameters adapted to residential electricity consumption can be adjusted first), ensuring that transformer operating status optimization accurately matches the core driving factors of load fluctuations and avoiding poor energy-saving effects or decreased power supply quality due to blind adjustments.

[0040] In one embodiment of this application, step S190, which determines the transformer operating state optimization parameters based on the real-time demand fluctuation vector using a fuzzy control algorithm, further includes the following steps: Figure 6 As shown, the specific content is as follows: Step S610: Based on the sliding time window algorithm, determine the demand fluctuation vector sequence according to the real-time demand fluctuation vector; Step S620: Determine the peak and valley characteristics based on the demand fluctuation vector sequence. The peak and valley characteristics include the peak load and the valley load. Step S630: Determine whether the peak load exceeds the preset peak threshold; Step S640: If it is determined that the load exceeds the preset peak threshold, then the transformer operating status optimization parameters are determined based on the fuzzy control algorithm according to the load peak. Step S650: If it is determined that the load peak value has not exceeded the preset peak value threshold, then determine whether the load valley value has exceeded the preset valley value threshold. Step S660: If it is determined that the preset valley value threshold has not been exceeded, the transformer operating status optimization parameters are determined based on the fuzzy control algorithm and the load valley value.

[0041] Specifically, a sliding time window algorithm can be used to set a fixed window length and step size, expanding a single instantaneous demand fluctuation vector into a continuous time-series demand fluctuation vector sequence. This reflects the continuous trend of load fluctuations and avoids misjudgments in parameter adjustment due to the instantaneous nature of a single vector. Peak-valley characteristics (including load peaks and load troughs) are identified using a first-order difference algorithm. The rate of change of the actual load at adjacent moments corresponding to the demand fluctuation vector sequence is calculated as: (Load at the next moment - Load at the previous moment) / Load at the previous moment × 100%. When the rate of change of two consecutive sampling points changes from positive to negative, the load at the previous moment is the load peak; when the rate of change changes from negative to positive, the load at the previous moment is the load trough. This ensures the accuracy of peak-valley identification. The fuzzy control algorithm is suitable for scenarios where there is no precise mathematical model for load fluctuations and parameter adjustments. By defining input fuzzy variables (load peak deviation), output fuzzy variables (transformer operating state optimization parameters), and a fuzzy rule base, a nonlinear mapping from load peaks to optimization parameters is achieved. Among them, the transformer operation status optimization parameters include the voltage regulation ratio (based on rated voltage) and the cooling system start threshold (temperature), which correspond to load stability and equipment protection requirements, respectively. In the valley scenario, fuzzy control takes energy saving as the core objective. The input fuzzy variable is the load valley deviation (preset valley threshold - load valley value), and the output fuzzy variables are the voltage reduction ratio and the start / stop command of redundant equipment. In this way, reducing voltage reduces iron loss, and shutting down redundant transformers reduces no-load loss.

[0042] For example, taking a residential community scenario, based on the sliding time window algorithm, a demand fluctuation vector sequence for the period 19:00-20:00 is generated according to the real-time demand fluctuation vector: [83kW positive fluctuation, 88kW positive fluctuation, 92kW positive fluctuation, 85kW positive fluctuation], corresponding to the actual load sequence [933kW, 938kW, 942kW, 935kW]. The adjacent load change rates are calculated using a first-order difference algorithm (+0.54% for 19:15-19:30, +0.43% for 19:30-19:45, and -0.74% for 19:45-20:00), identifying a load peak of 942kW (from 19:30-19:45, when the change rate turns from positive to negative) and a load trough of 933kW (from 19:00-19:15, the start point of the sequence). The load peak of 942kW is compared with a preset peak threshold of 940kW, indicating that it exceeds the preset peak threshold. Based on the fuzzy control algorithm, the load peak deviation of 2kW (load peak - preset peak threshold) is fuzzified into a small deviation. Combined with the fuzzy rule base (small deviation corresponds to voltage fine-tuning and medium cooling threshold), the defuzzified output transformer operation status optimization parameters are: voltage regulation ratio is 101% of the rated voltage of 380V (adjusted to 383.8V, ensuring voltage stability under high load), and cooling system start threshold is 74℃ (forced air cooling is started when the oil temperature exceeds 74℃). If the actual load corresponding to the demand fluctuation vector sequence is [895kW, 905kW, 910kW, 898kW], and the load peak of 910kW does not exceed the preset peak threshold of 940kW, then the load valley value of 895kW is further judged to be within the preset valley value threshold of 900kW (based on the low load energy saving threshold setting). It is found that it does not exceed the preset valley value threshold. Therefore, the load valley value deviation of 5kW (preset valley value threshold - load valley value) is fuzzified into a small deviation. According to the energy saving-oriented fuzzy rule, the optimization parameters are output: the voltage regulation ratio is 97% of the rated voltage (368.6V after adjustment, reducing iron loss), and one redundant transformer is shut down (reducing no-load loss by about 45%).

[0043] In the above method, by expanding a single instantaneous fluctuation vector into continuous time-series data, frequent parameter adjustments caused by misjudgment of instantaneous data are avoided, ensuring the continuity and stability of parameter optimization. Secondly, by identifying peak and valley features, the peak and valley values ​​of the load are accurately located, providing a clear basis for subsequent parameter adjustments based on different scenarios, solving the pain point that traditional methods cannot distinguish extreme values ​​of load fluctuations. Thirdly, by using hierarchical threshold judgment (preset peak threshold and preset valley threshold), differentiated goals of ensuring stability under high load and promoting energy saving under low load are achieved, avoiding the drawbacks of a one-size-fits-all parameter configuration. Finally, based on the fuzzy control algorithm, the uncertainty of load fluctuations (such as small deviations in peak and valley values) is adapted through nonlinear mapping, and the output optimized parameters such as voltage regulation ratio, cooling threshold, and redundant equipment instructions not only ensure power supply quality and equipment safety under high load, but also achieve precise energy saving under low load, completely solving the core problem that traditional static parameter configuration cannot adapt to dynamic load fluctuations, and providing an adaptive and highly reliable parameter optimization scheme for intelligent power saving of transformers.

[0044] In one embodiment of this application, after determining the transformer operating state optimization parameters based on the real-time demand fluctuation vector using the fuzzy control algorithm in step S190, the following steps are also included: Figure 7 As shown, the specific content is as follows: Step S710: Obtain the predicted and actual power load demand values ​​under different scenarios; Step S720: In any scenario, determine whether the weighted average absolute percentage error between the predicted power load demand and the actual power load demand exceeds a preset error threshold. Step S730: If the error exceeds the preset threshold, optimize the power load demand prediction model based on the Bayesian optimization algorithm.

[0045] Specifically, using the summer high-temperature period (with significant load fluctuations) from August 1st to August 3rd, 2024 as the time frame, the predicted and actual power load demand values ​​for two scenarios—residential communities and industrial parks—were obtained. In the residential community scenario, the predicted power load demand values ​​came from the random forest prediction model constructed in step S160 (e.g., predicted value of 920kW at 19:00 on August 1st, 935kW at 19:00 on August 2nd, and 940kW at 19:00 on August 3rd). The actual power load demand values ​​were collected every 15 minutes from the community's main smart meter and aggregated into hourly data (e.g., actual value of 1050kW at 19:00 on August 1st, 1080kW at 19:00 on August 2nd, and 1100kW at 19:00 on August 3rd). In the industrial park scenario, the predicted power load demand comes from an LSTM prediction model optimized for this scenario (e.g., a predicted value of 2750kW at 14:00 on August 1st), while the actual power load demand is collected hourly by the transformer load sensor of the industrial data acquisition system (e.g., an actual value of 2920kW at 14:00 on August 1st). Using the Weighted Mean Absolute Percentage Error (WMAPE) as the evaluation metric, and setting a preset error threshold of 8%, the WMAPE for the peak load period of 19:00-20:00 in a residential community scenario (weight set to 1.2) is calculated as follows: The numerator is |1050-920|×1.2+|1080-935|×1.2+|1100-940|×1.2=522, and the denominator is 1050×1.2+1080×1.2+1100×1.2=3876. The final WMAPE is 522 / 3876≈13.47%, exceeding the preset error threshold of 8%. In the industrial park scenario, the WMAPE calculated for the peak load period of 14:00-15:00 is approximately 9.2%, also exceeding the preset error threshold. Therefore, the power load demand prediction models for these two scenarios are optimized using a Bayesian optimization algorithm. For the random forest prediction model in residential communities, the range of hyperparameters to be optimized (number of decision trees, maximum depth, minimum number of split samples) was set to 400-800 trees, 10-25 layers, and 5-20 nodes. A probabilistic model was constructed using a Gaussian process and iterated 15 times using the expected improvement criterion. The optimal hyperparameters were finally determined to be 620 decision trees, 17 layers at maximum depth, and 9 nodes at minimum number of split samples. After retraining, the WMAPE decreased to 7.52%. For the LSTM prediction model in industrial parks, the number of hidden layer neurons was increased from 64 to 80, the learning rate was decreased from 0.001 to 0.0008, and the dropout rate was decreased from 0.2 to 0.15. After optimization, the WMAPE decreased to 7.8%.

[0046] In the above method, by collecting predicted and actual values ​​covering multiple scenarios such as residential communities and industrial parks, the optimization mechanism is ensured to be not limited to a single scenario and to adapt to the dynamic changes in different electricity consumption patterns (such as the surge in residential air conditioning load in summer and temporary production increases in industrial parks), thus avoiding the local optimum problem caused by traditional single-scenario optimization. Secondly, the weighted average absolute percentage error (WMAPE) is used as a quantitative indicator, and a preset error threshold is set to transform the subjective judgment of whether the model accuracy meets the standard into an objective data standard. Moreover, the weighting characteristics of WMAPE for high-load periods ensure that the error during critical scheduling periods is given priority, solving the pain point of traditional error indicators neglecting core periods. Finally, the prediction model is optimized using a Bayesian optimization algorithm, which improves efficiency compared to traditional grid search and can quickly locate the optimal hyperparameters (such as the number of decision trees in random forests and the learning rate of LSTM), effectively correcting the model accuracy decay caused by scenario changes (such as seasonal changes and production adjustments), ensuring that the power load demand prediction model maintains high accuracy in the long term, and guaranteeing the long-term effectiveness and stability of the transformer intelligent power saving solution.

[0047] In one embodiment of this application, in step S730, if it is determined that the preset error threshold is exceeded, after optimizing the power load demand prediction model based on the Bayesian optimization algorithm, the following steps are further included: Figure 8 As shown, the specific content is as follows: Step S810: Determine the sequence of dispatch instructions under different power loads based on the optimized power load demand forecasting model; Step S820: Determine the transformer power saving scheme based on the sequence of dispatch instructions under different power loads.

[0048] Specifically, different power loads can be divided into three levels: high load, medium load, and low load (the classification standard is based on the historical average load of the scenario and the rated capacity of the transformer), ensuring that the dispatch instruction sequence adapts to the differentiated needs of load fluctuations. The dispatch instruction sequence consists of specific operation instructions arranged in chronological order and load level, including voltage regulation ratio (based on rated voltage), cooling system start / stop threshold (oil temperature), and redundant transformer start / stop status, which respectively correspond to load stabilization, equipment protection, and energy saving and consumption reduction requirements.

[0049] For example, taking residential communities and industrial parks as examples, based on the optimized power load demand forecasting model, the power load is divided into three levels: high load, medium load, and low load, and a sequence of dispatch instructions is generated: In the residential community scenario, the transformer has a rated capacity of 1000kW and a rated voltage of 380V. High load (>900kW) corresponds to the evening peak period of 18:00-22:00, and the dispatch instructions are: voltage regulation ratio 101% (383.8V), cooling system start / stop threshold 74℃, and parallel operation of 2 transformers; medium load (700-900kW) corresponds to 8:00-11:00. During the 8:00 and 22:00-24:00 periods, the instructions are: voltage regulation ratio 100% (380V), cooling threshold 76℃, and 2 units in parallel. For low load (<700kW) during the 0:00-8:00 period, the instructions are: voltage regulation ratio 97% (368.6V), cooling threshold 80℃, and shutting down 1 redundant transformer. The final time-cycled scheduling instruction sequence is as follows: 0:00-8:00 low load instruction → 8:00-18:00 medium load instruction → 18:00-22:00 high load instruction → 22:00-24:00 medium load instruction. In an industrial park setting, the transformer has a rated capacity of 3000kW and a rated voltage of 10kV. High load (>2600kW) corresponds to peak production periods from 8:00-12:00 and 13:00-17:00, with instructions for a voltage regulation ratio of 100.5% (10050V), a cooling threshold of 72℃, and three units operating in parallel. Medium load (2000-2600kW) corresponds to the lunch break period from 12:00-13:00, with instructions for a voltage regulation ratio of 100% (10kV) and a cooling threshold of 7... 5℃, 2 units in parallel; Low load (<2000kW) corresponds to the shutdown period from 17:00 to 8:00 the next day. The instructions are: voltage regulation ratio 99.5% (9950V), cooling threshold 80℃, shut down 2 redundant transformers. The generated scheduling instruction sequence is: 0:00-8:00 low load instruction → 8:00-12:00 high load instruction → 12:00-13:00 medium load instruction → 13:00-17:00 high load instruction → 17:00-24:00 low load instruction. The scheduling instruction sequences of the above two scenarios are integrated into a transformer power saving scheme: The residential area scheme clarifies the applicable scenarios, transformer parameters and optimized model accuracy, refines the scheduling instruction details for each load period, and sets execution rules such as "collecting real-time load matching level every 15 minutes, voltage regulation is achieved through tap changers, and redundant transformers need to meet the load standard for 30 consecutive minutes before switching".The industrial park solution includes scenario information, load time period division, and instruction details. The execution rule is "load is collected every hour and the level is verified in conjunction with the production plan, and voltage regulation is achieved through on-load tap changers." This ultimately forms a differentiated energy-saving solution that can be directly implemented, achieving multi-scenario adaptation and avoiding the insufficient applicability of a single solution in different scenarios. This ensures the complete implementation of the intelligent energy-saving method for transformers and provides a replicable and scalable energy-saving solution for power grid energy conservation and consumption reduction.

[0050] This application also provides a transformer intelligent power-saving system based on multi-source data fusion. This system may include an acquisition module, a first determination module, a second determination module, a first judgment module, a second judgment module, a construction module, a third determination module, a fourth determination module, and a fifth determination module. The acquisition module acquires multi-source heterogeneous data, including meteorological data, industrial production data, and residential habit data. The first determination module determines a spatiotemporal correlation feature set based on the multi-source heterogeneous data. The second determination module determines the correlation coefficient between hourly fluctuation feature vectors and daily periodic change feature vectors based on the spatiotemporal correlation feature set using a long short-term memory network. The first judgment module determines whether the correlation coefficient exceeds a preset correlation coefficient threshold. If it exceeds the preset threshold, it acquires historical power load data and power load influencing factor data, and determines the relationship between historical power load data and power load influencing factors based on the Pearson correlation coefficient algorithm. The system comprises five modules: a first module for determining the coupling strength between factor data; a second module for determining whether the coupling strength exceeds a preset coupling strength threshold, and if so, for which the current power load influencing factor data is taken as a key influencing factor; a third module for constructing a power load demand forecasting model based on the random forest algorithm and the key influencing factors; a fourth module for inputting multi-source heterogeneous data into the power load demand forecasting model to output the power load demand forecast value; a fifth module for determining the instantaneous demand fluctuation vector based on the power load demand forecast value; and a sixth module for determining the transformer operating state optimization parameters based on the instantaneous demand fluctuation vector using a fuzzy control algorithm.

[0051] It should be noted that the embodiments of the transformer intelligent energy-saving system based on multi-source data fusion provided in this application can be used to execute the processing flow of the embodiments of the transformer intelligent energy-saving method based on multi-source data fusion in the above embodiments. Its functions will not be repeated here, but can be referred to the detailed description of the above method embodiments.

[0052] This application also provides an electronic device, which includes one or more processors and memory resources represented by a memory for storing instructions executable by the processor, such as application programs. The application programs stored in the memory may include one or more modules, each corresponding to a set of instructions. Furthermore, the processor is configured to execute instructions to perform the aforementioned intelligent power-saving method for transformers based on multi-source data fusion.

[0053] In one embodiment, a computer device, which may be a server, is also provided. The computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device stores data. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a transformer intelligent power-saving method based on multi-source data fusion.

[0054] In one embodiment, a computer device is provided, which may be a terminal. The computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a transformer intelligent power-saving method based on multi-source data fusion. The display unit of the computer device is used to form a visually visible image and may be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0055] It should be noted that although the steps of the transformer intelligent energy-saving method based on multi-source data fusion in this application are described in a specific order in the accompanying drawings, this does not require or imply that these steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps, such as omitting certain steps, combining multiple steps into one step, and / or decomposing one step into multiple steps, should all be considered part of this application.

[0056] It should be understood that this application is not limited to the detailed structure and arrangement of the modules in the intelligent transformer energy-saving system based on multi-source data fusion proposed in this specification. This application can have other implementations and can be implemented and executed in various ways. The foregoing variations and modifications fall within the scope of this application. It should be understood that the disclosure and definition of this application extend to all alternative combinations of two or more individual features mentioned or apparent in the text and / or drawings. All these different combinations constitute multiple alternative aspects of this application.

Claims

1. A smart power-saving method for transformers based on multi-source data fusion, characterized in that, include: Acquire multi-source heterogeneous data, including meteorological condition data, industrial production data, and resident habit data; Determine the spatiotemporal correlation feature set based on the multi-source heterogeneous data; Based on the Long Short-Term Memory Network, the correlation coefficient between the hourly fluctuation feature vector and the daily periodic change feature vector is determined according to the spatiotemporal correlation feature set. Determine whether the correlation coefficient exceeds a preset correlation coefficient threshold. If it exceeds the preset correlation coefficient threshold, obtain historical power load data and power load influencing factor data, and determine the coupling strength between the historical power load data and the power load influencing factor data based on the Pearson correlation coefficient algorithm. Determine whether the coupling strength exceeds a preset coupling strength threshold. If it is determined that the coupling strength exceeds the preset coupling strength threshold, then the current power load influencing factor data is taken as the key influencing factor. Based on the random forest algorithm, an electricity load demand prediction model is constructed according to the key influencing factors of electricity load. The multi-source heterogeneous data is input into the power load demand forecasting model to output the power load demand forecast value; Determine the instantaneous demand fluctuation vector based on the aforementioned electricity load demand forecast; Based on the fuzzy control algorithm, the optimal parameters for transformer operation status are determined according to the real-time demand fluctuation vector.

2. The intelligent power-saving method for transformers based on multi-source data fusion according to claim 1, characterized in that, The step of determining the spatiotemporal correlation feature set based on the multi-source heterogeneous data includes: Based on the K-means clustering algorithm, the multi-source heterogeneous data is clustered to determine the spatial partitioning characteristics of different spatial regions; Based on the sliding time window algorithm, time sequence features are extracted from the spatial partition features to determine the time window features of different time windows; Based on the Pearson correlation coefficient algorithm, the cross-latitudinal correlation coefficient matrix is ​​determined according to the time window characteristics; The spatiotemporal correlation feature set is determined based on the cross-latitudinal correlation coefficient matrix.

3. The intelligent power-saving method for transformers based on multi-source data fusion according to claim 1, characterized in that, The determination of the correlation coefficient between hourly fluctuation feature vectors and daily periodic change feature vectors based on the long short-term memory network and the spatiotemporal correlation feature set includes: Based on the sliding time window algorithm, hourly fluctuation feature vectors and daily periodic change feature vectors are obtained according to the spatiotemporal correlation feature set. A multi-dimensional spatiotemporal feature matrix is ​​determined based on the hourly fluctuation feature vector and the daily periodic change feature vector. Based on the Long Short-Term Memory Network, the coupling mode between the hourly fluctuation feature vector and the daily periodic change feature vector is determined according to the multi-dimensional spatiotemporal feature matrix. Based on the Pearson correlation coefficient algorithm, the correlation coefficient between the hourly fluctuation feature vector and the daily periodic change feature vector is determined according to the coupling mode.

4. The intelligent power-saving method for transformers based on multi-source data fusion according to claim 1, characterized in that, The electricity load demand forecasting model, based on the random forest algorithm and the key influencing factors of electricity load, includes: Based on the random forest algorithm, the key influencing factors of the power load are trained to determine the importance scores of the key influencing factors of the power load. Determine whether the importance score exceeds a preset score threshold; If the score exceeds the preset threshold, the current key factors affecting the power load will be used as prediction variables. Determine the set of predictor variables based on the predictor variables; The power load demand forecasting model is constructed based on the set of forecast variables.

5. The intelligent power-saving method for transformers based on multi-source data fusion according to claim 1, characterized in that, The step of determining the instantaneous demand fluctuation vector based on the predicted electricity load demand includes: Based on the predicted electricity load demand, meteorological fluctuation characteristics, industrial load fluctuation characteristics, and residential electricity consumption fluctuation characteristics are determined. The weights corresponding to the meteorological fluctuation characteristics, the industrial load fluctuation characteristics, and the residential electricity consumption fluctuation characteristics are obtained respectively. Based on the weights corresponding to the meteorological fluctuation characteristics, the industrial load fluctuation characteristics, and the residential electricity consumption fluctuation characteristics, a weighted calculation is performed on the meteorological fluctuation characteristics, the industrial load fluctuation characteristics, and the residential electricity consumption fluctuation characteristics, and the weighted calculation result is used as the real-time demand fluctuation vector.

6. The intelligent power-saving method for transformers based on multi-source data fusion according to claim 1, characterized in that, The method of determining transformer operating state optimization parameters based on the fuzzy control algorithm and the real-time demand fluctuation vector includes: Based on the sliding time window algorithm, a demand fluctuation vector sequence is determined according to the real-time demand fluctuation vector. Peak-valley characteristics are determined based on the demand fluctuation vector sequence, and the peak-valley characteristics include load peak values ​​and load trough values; Determine whether the load peak exceeds a preset peak threshold; If the load exceeds the preset peak threshold, the transformer operating state optimization parameters are determined based on the load peak value using a fuzzy control algorithm. If it is determined that the load trough value does not exceed the preset peak threshold, then it is determined whether the load trough value exceeds the preset trough value threshold. If it is determined that the load has not exceeded the preset valley threshold, the transformer operating status optimization parameters are determined based on the fuzzy control algorithm according to the load valley value.

7. The intelligent power-saving method for transformers based on multi-source data fusion according to claim 1, characterized in that, After determining the transformer operating state optimization parameters based on the real-time demand fluctuation vector using the fuzzy control algorithm, the process includes: Obtain the predicted and actual power load demand values ​​under different scenarios; In any scenario, determine whether the weighted average absolute percentage error between the predicted power load demand and the actual power load demand exceeds a preset error threshold. If the error exceeds the preset error threshold, the power load demand prediction model is optimized based on the Bayesian optimization algorithm.

8. The intelligent power-saving method for transformers based on multi-source data fusion according to claim 7, characterized in that, After optimizing the electricity load demand forecasting model based on the Bayesian optimization algorithm, the method further includes: Based on the optimized power load demand prediction model, determine the sequence of dispatch instructions under different power loads; The transformer energy-saving scheme is determined based on the sequence of dispatch instructions under different power loads.

9. A transformer intelligent energy-saving system based on multi-source data fusion, characterized in that, include: The acquisition module is used to acquire multi-source heterogeneous data, including meteorological condition data, industrial production data, and resident habit data. The first determining module is used to determine a spatiotemporal correlation feature set based on the multi-source heterogeneous data; The second determining module is used to determine the correlation coefficient between the hourly fluctuation feature vector and the daily periodic change feature vector based on the long short-term memory network and the spatiotemporal correlation feature set. The first judgment module is used to determine whether the correlation coefficient exceeds a preset correlation coefficient threshold. If it is determined that the correlation coefficient exceeds the preset correlation coefficient threshold, historical power load data and power load influencing factor data are obtained, and the coupling strength between the historical power load data and the power load influencing factor data is determined based on the Pearson correlation coefficient algorithm. The second judgment module is used to determine whether the coupling strength exceeds a preset coupling strength threshold. If it is determined that the coupling strength exceeds the preset coupling strength threshold, the current power load influencing factor data is taken as the key influencing factor. The module is used to build an electricity load demand prediction model based on the random forest algorithm and the key influencing factors of electricity load. The third determining module is used to input the multi-source heterogeneous data into the power load demand prediction model to output the power load demand prediction value. The fourth determining module is used to determine the instantaneous demand fluctuation vector based on the predicted power load demand value; The fifth determination module is used to determine the transformer operating state optimization parameters based on the real-time demand fluctuation vector using a fuzzy control algorithm.

Images