A method and device for overload risk alarm in power distribution substations based on multi-source sensing data fusion

Abstract

This invention discloses a method and device for overload risk alarm in substations based on multi-source sensing data fusion, belonging to the field of intelligent early warning technology. The method includes the following steps: during substation operation, a micro-amplitude synchronization pulse is injected during the multi-source sensing data acquisition phase; pulse feedback results are collected; the acquisition clock offset is calculated; a second-level alignment vector is constructed to achieve timing synchronization between different sensing channels; and a return-sensitive baseline is established. This invention achieves accurate early warning of abnormal load fluctuations by combining multi-source sensing data alignment and return-identification with edge fidelity filtering, heat accumulation calculation, and decoupling of fast and slow variables. It also establishes a closed-loop control mechanism based on residual feedback, significantly improving the accuracy and foresight of substation overload risk identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent early warning technology, specifically to a method and device for alarming overload risks in power distribution stations based on multi-source sensing data fusion. Background Technology

[0002] "Overload Risk Alarm for Substations Based on Multi-Source Sensing Data Fusion" refers to a system that, during the operation of a substation, no longer relies on data from a single monitoring point, but simultaneously collects sensing information from multiple sources, such as current, voltage, load power, conductor temperature, equipment status, and ambient temperature and humidity. This information from different dimensions is then processed and analyzed in a unified manner using data fusion technology, resulting in a comprehensive understanding of the substation's operating conditions. Based on this, the system can identify potential risks when the load gradually approaches or exceeds a safety threshold, and, by combining trend prediction and dynamic threshold adjustment, issue an overload risk alarm signal in advance. This method not only avoids misjudgments caused by single data anomalies but also improves the accuracy and real-time performance of overload risk identification, making substation operation safer and more reliable, and providing maintenance personnel with a basis for early intervention.

[0003] The existing technology has the following shortcomings:

[0004] In existing technologies, the prediction of substation overload risk largely relies on dynamic trend modeling methods based on historical load curves. However, when the load experiences severe reverse oscillations within a short period—for example, a sharp drop followed by a rapid rebound—existing prediction models often treat these anomalous fluctuations as noise or transient disturbances, and thus employ smoothing techniques to correct the prediction curve. Due to excessive smoothing, the model's output trend curve appears to show a low-risk or even stable trend, masking the actual, continuous increase in load. Once this problem occurs, the system will be unable to trigger an overload risk alarm in time before the critical state, potentially causing the substation load to exceed the safety threshold within minutes, rapidly amplifying the risk and escalating into an uncontrollable and severe accident.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for alarming overload risks in power distribution stations based on multi-source sensing data fusion, so as to solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a substation overload risk alarm method based on multi-source sensing data fusion, comprising the following steps:

[0008] S001, during the operation of the substation, injects a micro-amplitude synchronization pulse during the multi-source sensing data acquisition stage, collects the pulse feedback results, calculates the acquisition clock offset, constructs a second-level alignment vector, realizes the timing synchronization between different sensing channels, and establishes a foldback sensitive baseline;

[0009] S002, under the constraint of the return sensitive baseline, based on the second-level aligned multi-source sensing data, the continuous derivative sequence is calculated using the sliding window method, the fluctuation pattern of first falling and then rising is identified, the fluctuation segments exceeding the set amplitude threshold are screened, and the return window list is constructed.

[0010] S003, within the constraints of the foldback window list, performs edge fidelity filtering. By dynamically setting the filtering window range, it suppresses trend smoothing across foldback segments, preserves the waveform slope of the rising segment, eliminates peak clipping effects, and outputs enhanced trend data.

[0011] S004 After completing the edge fidelity filtering process, the temperature sensing data is frozen to decouple the change paths of fast and slow variables. The heat accumulation value is generated by the current square time accumulation calculation method, and the heat protection threshold is constructed to improve the overload risk level.

[0012] S005, after the overload risk level is increased, the weighting coefficients of the multi-source sensing channels are reallocated, the weight of current and power data is increased in the foldback window, the weight of temperature and environmental data is reduced, and a pre-warning index is generated as an alarm triggering lead signal.

[0013] S006 After obtaining the early warning index, the residual changes between the trend prediction data and the actual measurement data are compared, the threshold parameters, sliding window width and sensing channel weighting coefficients are dynamically updated, and the updated parameter set is written into the foldback sensitive baseline to complete the closed-loop risk alarm process from perception, identification, processing to control.

[0014] Preferably, step S001 includes:

[0015] The preset current sensor, voltage sensor, load detection device, conductor temperature sensor, equipment operation status acquisition device, and environmental temperature and humidity sensing device are capable of receiving external micro-amplitude synchronization pulses and recording timestamps;

[0016] In the initial stage of data acquisition, the central control unit sends a synchronization pulse signal with a voltage amplitude of less than 100 millivolts and a duration of less than 10 milliseconds to each sensing channel, and collects the receiving timestamps fed back by each channel and calculates the time difference.

[0017] By statistically analyzing the sampling clock offset pattern of each channel within multiple consecutive synchronization pulse cycles, a trend of time difference change per second is formed, and a second-level alignment vector is generated for unified time base correction.

[0018] The timestamps of all channel sampled data are adjusted based on the second-level alignment vector. The data time axis is corrected by time shifting or interpolation. A foldback sensitive baseline based on a unified standard time is constructed to provide time consistency assurance for subsequent abnormal fluctuation identification.

[0019] Preferably, step S002 includes:

[0020] Based on the time-aligned sensing data of the current channel, load power channel and conductor temperature channel, a sliding window is divided by sampling once per second, with a window width of 30 seconds and a step size of 1 second, and the continuous derivative sequence within each window is calculated.

[0021] Identify whether there is a fluctuation pattern formed by a combination of continuous falling segments and continuous rising segments in each window. Each segment change should have no less than 3 sampling points, and the amplitude should exceed 8% of the rated current value. The time interval between the preceding and following segments should be no less than 5 seconds.

[0022] For windows that initially meet the conditions, the load power channel and conductor temperature channel are retrieved and a consistency verification is performed within the same time interval to confirm that the power fluctuation direction is consistent and the amplitude exceeds 10% of the rated value, and the temperature channel rises by more than 1.5 degrees Celsius within 10 seconds after the end of the window.

[0023] The time periods that meet the multi-channel verification conditions are compiled into a structured list of return windows, and the start and end times of the windows, channel numbers, change ranges and verification results are recorded as input for subsequent trend retention and risk identification.

[0024] Preferably, step S003 includes:

[0025] Extract the start time, end time, channel type, and fluctuation characteristics of each return window, and extract the corresponding time period from the original sensing data and extend it by 1 second before and after as a protection area;

[0026] The sliding window width is dynamically set based on the maximum value of the derivative of the turnaround segment and the time span, and continuous filtering is performed using a weighted average method, with the weights gradually decreasing from the center of the window to both ends.

[0027] For continuous sampling points in the upward trend segment, a slope protection mechanism is enabled. When three sets of consecutive positive difference values ​​are met and the amplitude is not lower than the rated threshold, filtering is paused and the original data is retained.

[0028] Data that is detected as peak points are not modified. After all the turnaround segments are processed, the filtering results are embedded into the original data sequence, and linear interpolation is performed with a 2-second buffer before and after to ensure data continuity.

[0029] Preferably, step S004 includes:

[0030] Extract the sensing data sequence after edge fidelity filtering, acquire conductor temperature, equipment casing temperature and ambient temperature data within the corresponding time period of the foldback window, and keep the most recent effective temperature value before the start of the window unchanged within the window;

[0031] The current sensing data sampled per second within the foldback window is squared, and all squared current values ​​are accumulated to calculate the current square time product as the heat input.

[0032] By combining the equipment model, cable parameters and ventilation conditions to match the equipment's heat capacity carrying capacity, the product of the square of the current over time is compared with the heat capacity threshold to classify the overload risk levels as Level 1, Level 2 and Level 3.

[0033] The risk level, start and end time, peak current and freezing temperature corresponding to each turnaround window are written into the turnaround window list for use in subsequent dynamic weight adjustment and early warning index generation steps.

[0034] Preferably, step S005 includes:

[0035] Based on the risk level corresponding to the heat accumulation value within the return window, the weighting coefficient of the sensing channel is dynamically adjusted, increasing the weight of the current channel and power channel to no less than 1.5, and decreasing the weight of the conductor temperature channel and ambient temperature channel to no more than 0.6.

[0036] The current, power, conductor temperature and ambient temperature data at each sampling moment within the foldback window are multiplied by their corresponding adjusted weights, and the weighted values ​​are summed to generate a continuous weighted risk data sequence.

[0037] The normalization interval of each sensing channel is determined by using historical data within 24 hours. The weighted risk data sequence is standardized to generate a forward warning index distributed in the range of 0 to 100.

[0038] The early warning index is compared with the set risk classification threshold. If the index exceeds the preset threshold for three consecutive sampling periods, the current return window is marked as an abnormal and continuously developing state, and the weight configuration and index evolution trajectory are recorded.

[0039] Preferably, step S006 includes:

[0040] Calculate the residual per second between the trend prediction data and the actual perceived data within the loop window, generate a residual change sequence, and assess the degree of deviation of the prediction model;

[0041] The alarm parameters are adjusted based on the residual change results, including lowering the over-alarm thresholds for current and power, compressing the sliding window width, increasing the weight of fast variable channels and simultaneously decreasing the weight of slow variable channels.

[0042] Each parameter adjustment result is recorded as an independent parameter set, including the return window number, residual statistics, weights before and after the update, and changes in window width and alarm threshold;

[0043] The independent parameter set is written into the return sensitive baseline by timestamp and used as the priority configuration basis for the next round of risk assessment and perception preprocessing operations to achieve closed-loop adaptive control of the alarm process.

[0044] The substation overload risk alarm device based on multi-source sensing data fusion includes a data alignment module, a back-recognition module, a fidelity filtering module, a variable decoupling and risk modeling module, a weighted adjustment and early warning module, and a closed-loop control module.

[0045] The data alignment module injects micro-amplitude synchronization pulses during the multi-source sensing data acquisition stage in the operation of the substation, collects pulse feedback results, calculates the acquisition clock offset, constructs a second-level alignment vector, realizes timing synchronization between different sensing channels, and establishes a foldback sensitive baseline.

[0046] The foldback identification module, under the constraint of the foldback sensitive baseline, calculates the continuous derivative sequence based on the second-level aligned multi-source sensing data using a sliding window method, identifies the fluctuation pattern of first falling and then rising, filters fluctuation segments that exceed the set amplitude threshold, and constructs a foldback window list;

[0047] The fidelity filtering module performs edge fidelity filtering within the constraints of the foldback window list. By dynamically setting the filtering window range, it suppresses trend smoothing across foldback segments, preserves the waveform slope of the rising segment, eliminates peak clipping effects, and outputs enhanced trend data.

[0048] The variable decoupling and risk modeling module performs a freezing operation on temperature-related sensing data after completing edge-fidelity filtering, thereby decoupling the change paths of fast and slow variables. It generates heat accumulation values ​​by using the current square time cumulative calculation method, constructs a heat protection threshold, and improves the overload risk level.

[0049] The weighted adjustment and early warning module reallocates the weighting coefficients of the multi-source sensing channels after the overload risk level is increased. In the foldback window, the weight of current and power data is increased, while the weight of temperature and environmental data is decreased, and a pre-warning index is generated as an alarm triggering lead signal.

[0050] The closed-loop control module, after obtaining the early warning index, compares the residual changes between the trend prediction data and the actual measurement data, dynamically updates the threshold parameters, sliding window width and sensing channel weighting coefficients, and writes the updated parameter set into the foldback sensitive baseline, thus completing the closed-loop risk alarm process from perception, identification, processing to control.

[0051] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0052] This invention utilizes the temporal alignment and foldback feature identification of multiple types of sensing data to accurately capture abnormal patterns of sudden load drops and rises. It preserves key upward trends through edge-fidelity filtering, avoiding smoothing that masks risk signals. Simultaneously, it introduces a decoupling mechanism between heat accumulation and fast / slow variables to improve the accuracy of physical risk assessment. Furthermore, it establishes closed-loop adaptive control logic through early warning index generation and residual feedback correction, significantly enhancing the system's sensitivity to abnormal fluctuations and its forward-looking warning capabilities. Compared to existing technologies, this invention offers earlier warnings, faster responses, and more accurate risk identification, significantly improving the operational safety and risk management capabilities of substations under complex load dynamics. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0054] Figure 1 This is a flowchart of the overload risk alarm method for substations based on multi-source sensing data fusion according to the present invention.

[0055] Figure 2 This is a schematic diagram of the overload risk alarm device for power distribution stations based on multi-source sensing data fusion according to the present invention. Detailed Implementation

[0056] 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 the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0057] This invention provides, for example Figure 1 The overload risk alarm method for substations based on multi-source sensing data fusion, as shown, includes the following steps:

[0058] S001, during the operation of the substation, injects a micro-amplitude synchronization pulse during the multi-source sensing data acquisition stage, collects the pulse feedback results, calculates the acquisition clock offset, constructs a second-level alignment vector, realizes the timing synchronization between different sensing channels, and establishes a foldback sensitive baseline;

[0059] To achieve high-precision alignment of various sensing data within a substation in the time dimension, resolve the data fusion offset issue caused by asynchronous sampling clocks across different channels, and ensure accurate perception of short-cycle load changes and subsequent overload risk identification, the following specific steps are proposed:

[0060] During normal operation of the substation, for various sensor channels including current sensors, voltage sensors, load detection devices, conductor temperature sensors, equipment operation status acquisition devices, and environmental temperature and humidity sensing devices, each data acquisition channel is pre-configured to have hardware circuitry for receiving external pulse signals and the ability to record timestamps. In the initial stage of multi-channel data acquisition, a set of micro-amplitude synchronous voltage pulses is sent to all sensing channels at a preset period (e.g., 1 second) via physical connection lines or industrial wireless communication protocols (such as RS485, CAN, or wireless HART). The voltage amplitude of this pulse is controlled below the lower limit of the measurement range of each sensor, for example, less than 100 millivolts, and the pulse duration is set to less than 10 milliseconds to ensure no interference with the sensor data. This reference pulse signal is injected into the receiving pin at the front end of each channel, triggering the channel's local clock to record the receiving time, establishing a mapping relationship with the local sampling clock through an internal high-precision timer. After receiving the reference pulse signal, the sampling channel immediately generates a timestamp for the current time and feeds it back to the central control unit via the communication network.

[0061] After receiving the timestamps returned from all sensing channels, the central control unit selects the theoretical transmission time of each set of reference pulses as the standard time, compares it with the reception time points fed back by each channel, and calculates the time difference between each channel and the standard time. To ensure that this time difference reflects the offset behavior of the channel sampling clock itself, rather than accidental network delays or signal propagation jitter, the central control unit repeats this time offset measurement operation within multiple consecutive reference pulse cycles to obtain a stable time offset trend for each channel. This time offset trend reflects the systematic difference between the sampling clock inside the channel and the standard time reference, such as that caused by small deviations in crystal oscillator frequency, circuit response delays, and sampling start delays. Based on the above multi-cycle time offset measurement results, the central control unit performs second-by-second quantization processing on the actual clock offset value of each channel, forming a set of alignment vectors representing the difference between the channel and the unified standard time, which are used for subsequent data time reconstruction operations.

[0062] After acquiring the time alignment vectors for all channels, the raw sampled data from all sensing channels undergoes time axis correction. The central control unit adjusts the data timestamps for each channel based on the time offset in the aforementioned alignment vectors. Adjustment methods include interpolation, data shifting, or resampling. For sensing data with consistent frequencies, shifting can be used to move the original sampling points forward or backward by the corresponding time amount. For channels with non-uniform sampling or missing points, linear interpolation based on previous and subsequent sample values ​​can generate new sampling points with equal time intervals. The adjustment process strictly adheres to a unified standard time, ensuring that each corrected data set corresponds to the same actual physical time. The adjusted data is then reorganized according to a unified time series, forming a fully time-aligned sensing dataset. This process applies not only to high-frequency sensing data such as current and voltage signals but also covers parameters with lower update frequencies, such as conductor temperature and ambient temperature and humidity, thereby fusing them on a unified time base to form a cross-channel, cross-dimensional synchronous data sequence.

[0063] After completing the aforementioned time alignment and data reconstruction, the central control unit will use the adjusted data as a crucial basis for subsequent identification of short-period, drastic fluctuations and construct a return-sensitive baseline. The construction of the return-sensitive baseline is based on the alignment vector and the dynamic characteristics of the aligned data. By analyzing the time synchronization degree and change trend of each channel's data within a certain time sliding window, the synchronization disturbance characteristics of the data at a small time scale are determined. Whenever a channel experiences a significant increase in its rate of change within a specific time window, and its change time is time-off compared to other channels, the synchronization deviation evaluation index of the return-sensitive baseline is updated to determine whether the change is a real physical phenomenon or a pseudo-fluctuation caused by timing inconsistencies. The baseline serves to provide a timing reference for the subsequent return-sensitive feature identification process. Through continuous dynamic updates to this baseline, the system's accuracy in identifying short-term abnormal fluctuations can be improved, preventing sampling errors from interfering with fluctuation identification, and fundamentally solving the problem of misjudgment and missed judgment caused by multi-channel clock asynchrony in existing technologies.

[0064] This step aims to achieve high-precision time synchronization of data from multiple sensing channels in a substation, ensuring that real-time information collected by different types of sensors has a unified time reference, thereby eliminating data misalignment caused by sampling clock offsets. During substation operation, parameters such as current, voltage, load power, conductor temperature, and ambient temperature and humidity are often collected through different devices, sampling frequencies, and communication methods. Differences in sampling start points, sampling periods, and signal transmission delays between these channels can easily lead to time inconsistencies during data fusion analysis. This step, by injecting micro-amplitude synchronization pulses, establishing a time offset model, reconstructing data, and constructing a return-sensitive baseline, effectively corrects the time differences between channels, aligning all sensing data on a unified time axis. This provides a high-precision input foundation for subsequent return-fluctuation identification, trend tracking, and overload risk assessment. This approach avoids the errors caused by relying on single-channel timestamps or static smoothing methods in existing technologies, improving the accuracy of data fusion and the sensitivity of anomaly identification.

[0065] S002, under the constraint of the return sensitive baseline, based on the second-level aligned multi-source sensing data, the continuous derivative sequence is calculated using the sliding window method, the fluctuation pattern of first falling and then rising is identified, the fluctuation segments exceeding the set amplitude threshold are screened, and the return window list is constructed.

[0066] To accurately identify potentially severe short-cycle reverse fluctuations in substation operation based on time-aligned multi-source sensing data, and to effectively locate abnormal data segments with overload risks, thus supporting subsequent trend correction and risk warning processing, a technical method for constructing a turnaround window list is proposed. This method uses second-level synchronized multi-channel sensing data to identify continuous fluctuation intervals that conform to the "first decrease, then increase" characteristic, and its reliability is confirmed through multi-dimensional verification. The specific implementation is as follows:

[0067] After completing the time alignment of the substation sensing data channels, key sensing indicators were selected for fluctuation identification. The selected sensing channels include a current channel (reflecting actual power transmission intensity), a load power channel (characterizing changes in user-side energy demand), and a conductor temperature channel (tracking heat load conduction response). The sampling frequency is 1Hz, meaning one data point is collected per second. These sensing data are divided into sliding windows along a unified time axis. Each window contains 30 consecutive sampling points, with a window width of 30 seconds and a sliding step of 1 second. Within each window, the differences between adjacent sampling points are calculated sequentially according to the time series, forming a derivative sequence containing 29 difference values. This derivative sequence is used to evaluate the direction and rate of change of the data within that window. A negative difference indicates a downward trend, while a positive difference indicates an upward trend. The transition from consecutive negative values ​​to consecutive positive values ​​may constitute the start and end points of a zigzag behavior.

[0068] In the obtained derivative sequence, the system analyzes whether a clear fluctuation structure exists within the window, i.e., a continuous downward process followed by a continuous upward process, with the overall change exceeding a preset threshold. Taking the current channel as an example, the judgment criteria are set as follows: the downward segment must contain at least three consecutive negative values, and the maximum downward amplitude must be no less than 8% of the current rated current; subsequently, at least three consecutive positive values ​​must appear, and the maximum upward amplitude must also be no less than 8% of the current rated current. Furthermore, to ensure that the identified fluctuation segment does not belong to sensor measurement errors or instantaneous disturbances, the time interval between the downward and upward segments must be no less than 5 seconds, i.e., it must be a structurally continuous reverse change behavior. Windows that meet the above judgment criteria are initially identified as suspected zigzag windows, and their start time, end time, and maximum change amplitude are marked and stored in the internal structure.

[0069] To improve identification accuracy and eliminate single-channel misjudgments, multi-channel verification was conducted on the time period initially identified as a return window. Using this window's time interval as a benchmark, data sequences from the load power channel and conductor temperature channel within the same time range were retrieved, and the same derivative structure analysis was performed on each. In the load power channel, the presence of a return fluctuation structure in the same direction as the current channel was checked. In the conductor temperature channel, a sustained temperature rise was checked for a 3-10 second period after the return window ended, indicating a genuine thermal load response. If fluctuation behavior consistent with the current channel structure was found in both auxiliary channels, and the magnitude of the change exceeded their respective preset thresholds (e.g., power change exceeding 10% of rated power, temperature rise exceeding 1.5 degrees Celsius), the return window was confirmed to be physically plausible and moved from the candidate list to the formal return window set.

[0070] All verified and confirmed retracement windows are compiled and organized to construct a complete list of retracement windows. This list stores the time period information of each retracement window in a structured format, including start time, end time, channel number, descent amplitude, rise amplitude, maximum derivative value, and matching information of relevant auxiliary channels. This list will serve as a boundary reference in subsequent processing to prevent data from being misidentified as transient noise and smoothed out during the smoothing and filtering process. In particular, before processing the dynamic trend prediction model, this list will serve as an input control parameter, indicating the structural regions that the prediction model should retain, thereby improving the accuracy of trend identification and providing a traceable foundation for subsequent risk level assessment and dynamic adjustment of early warning thresholds.

[0071] This step aims to accurately identify abnormal fluctuations—characterized by a short-cycle decline followed by a rise—from multi-source, time-aligned sensing data during substation operation. It extracts physically significant fluctuation segments and constructs a list of return windows for subsequent trend correction and risk warning processing. In actual operation, key indicators such as current and load power may experience severe reverse oscillations due to sudden load changes within a very short time. If such behavior is not identified in time, it is easily misjudged as noise by subsequent smoothing processes, masking the true risk signal. This step ensures that each extracted return fluctuation has temporal continuity, significant amplitude, and multi-channel consistency through sliding window division, derivative sequence calculation, fluctuation structure judgment, and cross-channel verification, providing a real physical background and operational basis. The constructed list of return windows serves as a benchmark for retaining key fluctuation structures, preventing erroneous smoothing from causing deviations in system risk assessment. It provides high-precision input boundaries for dynamic filtering strategies, threshold adjustments, and early warning index generation, significantly improving the real-time performance and accuracy of substation overload risk identification.

[0072] S003, within the constraints of the foldback window list, performs edge fidelity filtering. By dynamically setting the filtering window range, it suppresses trend smoothing across foldback segments, preserves the waveform slope of the rising segment, eliminates peak clipping effects, and outputs enhanced trend data.

[0073] To maximize the preservation of the true trend characteristics of the data during the identified zigzag fluctuations and avoid the weakening or masking of key features by traditional smoothing methods, thereby enhancing the authenticity of trend data and the accuracy of subsequent risk assessment, this step proposes an edge-fidelity filtering method based on a zigzag window list for targeted filtering correction of zigzag segments in sensing channels such as load, current, and power. This method, aiming to achieve data continuity and readability without structural damage, proposes a complete processing flow combining boundary constraints, slope preservation, and transition compensation. The specific implementation is as follows:

[0074] Based on the list of foldback windows constructed in the previous step, the start time, end time, channel type involved, maximum fluctuation amplitude, and derivative change rate of each foldback record are extracted. Taking the current sensing channel as an example, assume that the start time of a certain foldback window record is 08:12:30 and the end time is 08:12:58, with a time span of 28 seconds. According to this time period, 29 consecutive sampling points from 08:12:30 to 08:12:58 are extracted from the raw sampling data of the current channel, with a sampling frequency of 1Hz, i.e., one data point per second. To prevent the loss of edge information, a 1-second extension is added at both ends of the window as a protection zone, ultimately forming a complete processing segment containing 31 points. Each data segment will undergo independent filtering operations in subsequent processing to ensure that the window features are not masked by the overall trend.

[0075] The filtering window parameters are dynamically set based on the duration of each foldback window and the data fluctuation characteristics. If the fluctuation rate of a certain foldback segment is high (e.g., the maximum derivative value reaches more than 15% of the rated current), the filtering window width is set to 5 seconds, i.e., 5 sampling points are processed each time; if the fluctuation rate is low, but the time span exceeds 30 seconds, the filtering window width is set to more than 10 seconds to improve the overall trend smoothing capability. In actual operation, the filtering process is carried out in a time series manner, processing the data segments within each window sequentially from left to right. At each position of the sliding window, the 5 sampling points are processed by weighted averaging. The weighting factor is set according to the position of the data in the window, decreasing from the center to both ends, with the center point having the largest weight and the edge points having the smallest weight, to enhance the response capability to the center segment of fluctuation and weaken the influence of edge data.

[0076] During the sliding processing, a slope protection mechanism is employed for continuous data points located on the upward slope of the turnaround section. Under this mechanism, before each sliding calculation, the upward trend of the data within that section is assessed for its significance. The criterion is that three consecutive sets of differences are positive, and each set of differences is not less than a set threshold, such as an increase of not less than 2% of the rated current per second. When the above conditions are met, the weighted average processing of that section is paused, and the original data points are directly retained as part of the trend output. This mechanism effectively prevents the traditional sliding processing method from flattening the slope and avoids the upward trend being mistakenly weakened. Simultaneously, when a peak point is detected within the window (i.e., two consecutive leading data points with positive derivatives and two subsequent data points with negative derivatives), the original data remains unchanged, and no processing is allowed on that peak point, thus preventing the trend peak from being truncated.

[0077] After processing each turnaround window independently, each filtered trend data segment is embedded back into the original sensed data sequence, precisely replacing its corresponding time period. To eliminate potential numerical abrupt changes or discontinuities at data splicing points, a transition buffer with a transition length of 2 seconds is set at both ends of each turnaround segment. The connectivity value of the transition segment is calculated using linear interpolation, ensuring a natural and continuous connection between the processed segment and the context data. This ultimately forms a set of trend data that retains true structural characteristics, suppresses invalid high-frequency jitter, and possesses good smoothness and temporal continuity, providing an accurate data foundation for subsequent early warning index generation and risk level assessment.

[0078] This step, after identifying and locating the zigzag fluctuations, performs fidelity filtering with boundary control capabilities on these abnormal fluctuation intervals to preserve the key upward trend slope and peak shape, preventing trend information from being mistakenly weakened or clipped during data smoothing. Traditional moving average or median filtering methods often smooth out data edges and suppress peak values ​​when dealing with sharp fluctuations, masking abnormal trends in key indicators such as load and current, thus affecting the accuracy of overload risk identification. To address this issue, this step sets an independent filtering range for each abnormal fluctuation interval based on the time boundary of the zigzag window and dynamically adjusts the size and weighting method of the moving window. Simultaneously, a slope protection mechanism and peak preservation strategy are introduced to ensure that data is not misprocessed during rapid upward phases. The filtered data is seamlessly stitched with context data through transition compensation to form a complete perception sequence with enhanced trend. This step provides a structurally accurate and trend-clear data foundation for subsequent risk level enhancement and early warning indicator extraction, significantly improving the ability to perceive abnormal situations while maintaining data smoothness.

[0079] S004 After completing the edge fidelity filtering process, the temperature sensing data is frozen to decouple the change paths of fast and slow variables. The heat accumulation value is generated by the current square time accumulation calculation method, and the heat protection threshold is constructed to improve the overload risk level.

[0080] To improve the accuracy of risk identification after edge-fidelity filtering and avoid misjudging the risk level during the turnaround period due to the lag in temperature data response, a dynamic risk enhancement method based on the current thermal effect and decoupling of fast and slow variables is proposed in the multi-source sensing data processing flow of the substation. This method freezes lag-type variables, extracts the intensity accumulation features of fast variables, and constructs a heat protection threshold based on the actual equipment operating capacity, thereby achieving accurate and dynamic enhancement of the potential risk level within the turnaround window. The specific implementation is as follows:

[0081] After completing the preceding edge-fidelity filtering process, the corrected sensing data sequence output from this step is acquired. The focus is on extracting conductor temperature, equipment casing temperature, and ambient temperature data within the time period corresponding to the foldback window. These temperature data are typically acquired by thermistors, platinum resistance thermometers, or thermocouples. Their response speed is affected by heat capacity, dielectric conductivity, and measurement loop hysteresis, generally lagging behind the dynamic response of fast variables such as current and voltage. In practical engineering, even if the current fluctuates drastically, the temperature data often shows a slow rise or no significant change. To prevent this hysteresis from interfering with risk identification, the temperature sensing data within the time period corresponding to the foldback window is frozen. This means acquiring the most recent valid temperature value before the start of the window, maintaining this temperature value within the window, and restoring the normal acquisition value at the end of the window. This ensures that the temperature input remains stable within this time period, avoiding misjudgments of a low-risk state due to delayed reflection of true risk by lagging data.

[0082] While the temperature data is frozen, the corresponding current sensing data within the foldback window is processed for thermal intensity calculation. Specifically, the sampling value per second of the current channel is selected, and a square operation is performed on each data point to obtain the instantaneous thermal contribution value of the current intensity per second. Subsequently, the squared current values ​​within the entire foldback window are accumulated item by item to form the current square-time product value within that window. Taking actual power distribution equipment as an example, if the window length is 28 seconds, 28 current squared values ​​will be generated within the window. The accumulated total value represents the approximate heat input of the current to the conductor and the internal structure of the equipment during that time period. This value is unaffected by the lag of temperature changes and can reflect the energy load level borne by the equipment in real time. This calculation method follows Joule's law, which states that the heat generated by current passing through a conductor is proportional to the square of the current and the time, making it a reliable indicator for assessing the thermal stress of equipment.

[0083] After obtaining the current squared-time product, the corresponding heat capacity is matched and a heat protection threshold is constructed by referring to engineering design parameters such as equipment type, cable cross-sectional area, conductor material, and ventilation conditions. Taking a 10 kV switchgear as an example, a heat threshold upper limit is set based on its rated continuous operating capacity and short-term load capacity. For example, a certain type of switchgear is allowed to accumulate heat input within 5 minutes without exceeding a certain value. The heat product calculated in the current foldback window is compared with this upper limit value, and risk level intervals are divided proportionally: when the heat product exceeds 70% but is less than 90% of the allowable upper limit, the risk level is judged as "Level 1 Warning"; when the heat product reaches between 90% and 100%, the risk level is judged as "Level 2 Warning"; when the heat product exceeds 100%, that is, exceeds the allowable heat capacity range of the equipment, it is judged as "Level 3 Warning" and is considered to have a significant overload risk. In this way, the system's response to load fluctuations at the thermal effect level has a quantitative standard and classification basis, avoiding rough judgments based solely on instantaneous current values ​​or temperature hysteresis data.

[0084] After completing the heat product calculation and risk level determination, the corresponding risk level result is recorded in the turnaround window list, along with the start and end times, peak current, and temperature freeze value of that window. Simultaneously, this risk level is passed to subsequent weight adjustment and early warning index generation steps to increase the reference weight of current-related indicators during the turnaround period and reduce the proportion of temperature-related data in risk assessment, forming a dynamic weighting mechanism. The system can also combine the risk level sequence of multiple consecutive turnaround windows to determine if there is a cumulative overload trend. If consecutive high-level risks occur, an early warning signal can be further issued to the upper-level operation and maintenance terminal. Through the above process, a complete processing flow is achieved from structural trend fidelity, lag data freezing, thermal effect measurement to risk level quantification and improvement, ensuring that potential overload risks can be identified and responded to in advance under conditions of severe load fluctuations.

[0085] This step aims to improve the timeliness and accuracy of overload risk assessment during the turnaround window, addressing the issues of response lag and delayed warnings inherent in traditional temperature threshold-based assessments. Because temperature-sensing data typically exhibits strong inertia and thermal capacity effects, it cannot promptly reflect drastic changes in current or load over a short period, potentially leading to misjudgments of low risk even when an actual overload is imminent. Therefore, this step freezes temperature data within the turnaround window to block its lag effect and introduces a heat assessment method based on the current square time product. By calculating the total heat load borne by the equipment during this time period, a real-time quantitative assessment of the equipment's thermal stress is generated. Subsequently, combining the equipment's structural parameters and thermal capacity, a heat protection threshold is constructed, and risk levels are categorized accordingly, enabling a dynamic escalation of risk from "normal" to "warning" and then to "severe." This mechanism not only decouples fast and slow variables but also introduces physical-level risk assessment criteria, effectively improving the system's response speed to overload trends and providing a scientific and real-time basis for subsequent warning decisions.

[0086] S005, after the overload risk level is increased, the weighting coefficients of the multi-source sensing channels are reallocated, the weight of current and power data is increased in the foldback window, the weight of temperature and environmental data is reduced, and a pre-warning index is generated as an alarm triggering lead signal.

[0087] To improve the accuracy of overload risk perception during the turnaround window and to identify latent abnormal states that may evolve into faults in advance, after completing the risk level enhancement operation based on heat product, the risk discrimination contribution of various sensing data needs to be dynamically reconstructed to generate a forward warning index with predictive and trend response capabilities. By weighting current, power, temperature, and environmental parameters, the response capability of fast variables is enhanced while the hysteresis interference of slow variables is weakened, effectively improving the foresight and accuracy of the overall alarm system. The specific implementation method is as follows:

[0088] In the previous step, the heat product within the turnaround window was calculated, and this heat value was compared with the corresponding equipment heat capacity threshold to determine the current risk level. Based on the risk level result, the dynamic weights of the sensing channels involved in the current turnaround window are immediately adjusted. Taking a 10 kV switchgear in a substation as an example, its corresponding sensing channels include the current channel (sampling frequency of 1 Hz, unit: amperes), the active power channel (unit: kilowatts), the conductor temperature channel (unit: degrees Celsius), and the ambient temperature channel (unit: degrees Celsius). When the risk level is at Level 1 warning, the weight of the current channel is increased from the base value of 1.0 to 1.8, the weight of the power channel is increased from 1.0 to 1.5, the weight of the conductor temperature channel is decreased from 1.0 to 0.6, and the weight of the ambient temperature channel is decreased from 1.0 to 0.5. If the risk level is further increased to Level 2 or Level 3 warning, the weight of fast variables is increased proportionally until it reaches 2.0, while the proportion of slow variable channels is reduced simultaneously to ensure that the rapid response indicators dominate during the critical period of fluctuation.

[0089] Based on the adjusted channel weights, the sensing data at each moment in the foldback window is weighted. The operation process is as follows: starting from the beginning of the window, the corresponding sampled value of each channel is read second by second; the current value is multiplied by the adjusted weight of the current channel, the power value is multiplied by the adjusted weight of the power channel, and the conductor temperature value and ambient temperature value are multiplied by their respective reduced weight coefficients. Taking a certain moment as an example, if the sampled values ​​are 220 amperes current, 65 kilowatts power, 56 degrees Celsius conductor temperature, and 40 degrees Celsius ambient temperature, with corresponding weights of 2.0, 1.5, 0.6, and 0.5 respectively, then the weighted data are: current weighted value 440, power weighted value 97.5, conductor temperature weighted value 33.6, and ambient temperature weighted value 20. The four weighted values ​​are superimposed to obtain a total instantaneous risk impact value, which is used to characterize the intensity of the abnormal features at that moment. This processing operation is applied to every sampling moment of the entire turnaround window, forming a continuous weighted risk data sequence that fully describes the risk response change process after the perception parameter-dominated structural adjustment within the turnaround window.

[0090] After obtaining the complete risk data sequence, it is standardized and a pre-warning index is constructed. The standardization steps are as follows: First, based on the historical data of each channel over the past 24 hours, the historical maximum and minimum values ​​of current, power, and temperature data are calculated to construct a normalized interval; then, the aforementioned weighted data sequence is converted according to this normalized interval, compressed into an index range of 0 to 100. The pre-warning index constructed in this way not only has a unified numerical representation capability but also accommodates differences in numerical dimensions across multiple channels, improving its practical usability for interpretation. Furthermore, to enhance the operability of the index in early warning applications, the index range is further divided into risk zones. For example, 0 to 30 indicates stable operation, 30 to 60 indicates an upward trend in risk, 60 to 80 indicates entering a warning state, and above 80 is considered a high-risk state requiring an alarm decision. This pre-warning index is a trend expression formed based on enhanced processing of risk-sensitive channels during the turnaround window, providing a basis for early warning before a fault actually occurs.

[0091] The early warning index is fed back into the alarm logic of the current processing cycle and serves as a preliminary basis for subsequent alarm triggering. In the control flow, if the early warning index exceeds the preset warning threshold (e.g., 60) for three consecutive sampling cycles and remains within the turnaround window time range, the system automatically marks the current window as an abnormal, continuously developing state and records the weight adjustment configuration and the evolution trajectory of the early warning index during that period for subsequent closed-loop adjustment and alarm strategy updates. Simultaneously, if the current index rises rapidly but temperature-related channels show no significant response, a further heat accumulation prediction process can be initiated to confirm whether the current load has reached its heat capacity limit. Throughout this process, the early warning index not only provides a basis for judging alarm precursors but also becomes an important output result of the dynamic evolution of the dominant force in the risk modeling process.

[0092] The purpose of this step is to adjust the participation intensity of different types of sensing channels in risk assessment in a timely manner after the overload risk level increases during the turnaround window. By dynamically allocating weights, it enhances the dominance of fast variables such as current and power on the risk index, while weakening the lagging effects of slow variables such as temperature and environment, thereby constructing a more real-time and predictive early warning index. In substation operation, traditional alarm mechanisms often rely on fixed thresholds and static indicators, making it difficult to identify abnormal fluctuation trends within short periods, especially when load surges occur but temperature has not yet responded, leading to missed detections. Through the dynamic reconstruction of the weighted structure of sensing data in this step, the system can proactively increase the sensitivity of its response to key indicators based on the real-time risk level and construct a unified quantitative early warning index, providing a leading signal for the alarm logic. This index features continuous tracking, hierarchical management, and sensitive response, helping to detect abnormal situations that may evolve into overload in advance, improving the initiative and reliability of risk identification and early warning decisions throughout the substation.

[0093] S006 After obtaining the early warning index, the residual changes between the trend prediction data and the actual measurement data are compared, the threshold parameters, sliding window width and sensing channel weighting coefficients are dynamically updated, and the updated parameter set is written into the foldback sensitive baseline to complete the closed-loop risk alarm process from perception, identification, processing to control.

[0094] To construct an overload risk alarm process with closed-loop control capabilities, after obtaining the early warning index, it is necessary to continuously evaluate the deviation between the prediction model output and the actual observation results by combining the index with the operational trajectory of actual sensing data. This is achieved by analyzing the residual changes over time periods and dynamically updating relevant risk judgment parameters, including alarm thresholds, sliding window widths, and sensing channel weighting coefficients. Finally, the updated results are written into the foldback sensitive baseline to complete the closed-loop chain from sensing, identification, processing to control. The specific implementation method is as follows:

[0095] During the turnaround window, the system generates a corresponding advance warning index sequence and records its change trajectory at the second level. This warning index is obtained by weighted reconstruction of the sensing channels such as current, active power, conductor temperature, and ambient temperature in the aforementioned steps, and its value reflects the intensity of abnormal aggregation of the overall sensing data during this period. To determine the effectiveness of the warning index, it is compared with the actual sensing data during this period. Taking the current channel as an example, the trend prediction curve of the current turnaround window is extracted, which is the hourly upward trend reference value generated by the sliding window method; then, the measured current data of the same time period is extracted, and the difference between the prediction curve and the measured value per second is calculated to form a residual change sequence. This residual sequence is used to quantify the deviation between the model expectation and the actual operation. If the residual is found to be positive for a long time, it indicates that the system underestimates the rate of load increase; if the residual changes frequently and fluctuates drastically, it indicates that the model's response to fluctuations is unstable, and there is a risk of false alarms or missed alarms.

[0096] Based on the residual analysis results above, an adaptive update process for risk assessment parameters is initiated. First, the alarm thresholds are adjusted: if the average residual value is greater than zero, it indicates that the trend prediction is generally low. In this case, the current alarm threshold is adjusted from the original 10% over-limit to 8% over-limit; the power threshold is adjusted from 15% to 12%. Second, the sliding window width is adjusted: if the residual peak change rate exceeds the set stability threshold, for example, a change of more than 20% within 5 seconds, it indicates that the window width is too wide. The original 30-second sliding window needs to be compressed to 20 seconds to promptly capture short-cycle fluctuation characteristics. Third, the sensing channel weight configuration is adjusted: if the residual values ​​of the current and power channels account for more than 70% within this time period, while the temperature and environmental channel changes lag behind, the current channel weight is increased from 2.0 to 2.2, the power channel weight is increased from 1.8 to 2.0, and the conductor temperature channel weight is decreased from 0.6 to 0.5, and the environmental temperature weight is decreased from 0.5 to 0.4. By updating the parameters mentioned above, the model's response strength to risks driven by fast variables can be improved, while the risk of misjudgment caused by the lag of slow variables can be reduced.

[0097] A parameter set containing all updated content is constructed and written into the return-sensitive baseline structure. This parameter set includes fields such as the current return window number, corresponding start and end times, alarm threshold correction records, sliding window adjustment records, perception channel weight configuration, residual sequence statistics, and the maximum value of the pre-warning index that triggered this update. The writing method is as follows: a new parameter entry is inserted into the time series structure of the return-sensitive baseline, and it is marked with the timestamp of the update source, the associated risk level change process, and the subsequent effective scope. The baseline structure is a dynamically expandable data structure, which can form independent parameter records based on the feedback within each window, without affecting previous historical judgments or causing confusion in later models. When similar return structures reappear in the future, this parameter set can serve as an adaptive configuration basis, improving the prediction accuracy of the model under similar risk evolution conditions.

[0098] Based on the latest return-sensitive baseline after writing, the next round of return window monitoring and risk assessment process is reinitialized. When starting the next data processing loop, the system will prioritize calling the latest parameter configuration for sensing data preprocessing and risk assessment, ensuring that the model structure remains highly consistent with the current actual operating state. If high residuals continue to occur in subsequent runs, the system can repeat the current process for dynamic adjustment until the model stabilizes and converges. This closed-loop process has three characteristics: self-driven, self-feedback, and adaptive, solving the problem in existing technologies where model parameters rely on manual static configuration and cannot self-adjust based on operating results.

[0099] This step aims to achieve closed-loop adaptive control of the substation overload risk identification mechanism, ensuring that after handling foldback fluctuations and generating early warning indices, the system can continuously correct and optimize judgment parameters based on actual operating data. By analyzing the residuals between trend prediction data and actual sensing data in real time, the degree of deviation in the model's risk prediction is identified, and alarm thresholds, sliding window widths, and weighted configurations of sensing channels are dynamically adjusted accordingly, making the model more closely aligned with current operating conditions. Writing the updated parameters to the foldback sensitive baseline not only records the risk identification strategy historically but also provides an automatic adaptive reference for subsequent similar fluctuation scenarios. Compared to traditional static parameter setting methods, this method possesses significant dynamic response and continuous optimization capabilities, effectively improving the accuracy and timeliness of risk warnings, avoiding false alarms, missed alarms, or response delays caused by model rigidity, and constructing a closed-loop risk mechanism covering the entire process from data sensing to model control.

[0100] This invention utilizes the temporal alignment and foldback feature identification of multiple types of sensing data to accurately capture abnormal patterns of sudden load drops and rises. It preserves key upward trends through edge-fidelity filtering, avoiding smoothing that masks risk signals. Simultaneously, it introduces a decoupling mechanism between heat accumulation and fast / slow variables to improve the accuracy of physical risk assessment. Furthermore, it establishes closed-loop adaptive control logic through early warning index generation and residual feedback correction, significantly enhancing the system's sensitivity to abnormal fluctuations and its forward-looking warning capabilities. Compared to existing technologies, this invention offers earlier warnings, faster responses, and more accurate risk identification, significantly improving the operational safety and risk management capabilities of substations under complex load dynamics.

[0101] This invention provides, for example Figure 2 The substation overload risk alarm device based on multi-source sensing data fusion shown includes a data alignment module, a feedback identification module, a fidelity filtering module, a variable decoupling and risk modeling module, a weighted adjustment and early warning module, and a closed-loop control module.

[0102] The data alignment module injects micro-amplitude synchronization pulses during the multi-source sensing data acquisition stage in the operation of the substation, collects pulse feedback results, calculates the acquisition clock offset, constructs a second-level alignment vector, realizes timing synchronization between different sensing channels, and establishes a foldback sensitive baseline.

[0103] The foldback identification module, under the constraint of the foldback sensitive baseline, calculates the continuous derivative sequence based on the second-level aligned multi-source sensing data using a sliding window method, identifies the fluctuation pattern of first falling and then rising, filters fluctuation segments that exceed the set amplitude threshold, and constructs a foldback window list;

[0104] The fidelity filtering module performs edge fidelity filtering within the constraints of the foldback window list. By dynamically setting the filtering window range, it suppresses trend smoothing across foldback segments, preserves the waveform slope of the rising segment, eliminates peak clipping effects, and outputs enhanced trend data.

[0105] The variable decoupling and risk modeling module performs a freezing operation on temperature-related sensing data after completing edge-fidelity filtering, thereby decoupling the change paths of fast and slow variables. It generates heat accumulation values ​​by using the current square time cumulative calculation method, constructs a heat protection threshold, and improves the overload risk level.

[0106] The weighted adjustment and early warning module reallocates the weighting coefficients of the multi-source sensing channels after the overload risk level is increased. In the foldback window, the weight of current and power data is increased, while the weight of temperature and environmental data is decreased, and a pre-warning index is generated as an alarm triggering lead signal.

[0107] The closed-loop control module, after obtaining the early warning index, compares the residual changes between the trend prediction data and the actual measurement data, dynamically updates the threshold parameters, sliding window width and sensing channel weighting coefficients, and writes the updated parameter set into the foldback sensitive baseline, thus completing the closed-loop risk alarm process from perception, identification, processing to control.

[0108] The overload risk alarm method for substations based on multi-source sensing data fusion provided in this invention is implemented through the aforementioned overload risk alarm device for substations based on multi-source sensing data fusion. For details on the specific methods and processes of the overload risk alarm device for substations based on multi-source sensing data fusion, please refer to the embodiments of the overload risk alarm method for substations based on multi-source sensing data fusion, which will not be repeated here.

[0109] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for overload risk alarm in substations based on multi-source sensing data fusion, characterized in that, Includes the following steps: S001, during the operation of the substation, injects a micro-amplitude synchronization pulse during the multi-source sensing data acquisition stage, collects the pulse feedback results, calculates the acquisition clock offset, constructs a second-level alignment vector, realizes the timing synchronization between different sensing channels, and establishes a foldback sensitive baseline; S002, under the constraint of the return sensitive baseline, based on the second-level aligned multi-source sensing data, the continuous derivative sequence is calculated using the sliding window method, the fluctuation pattern of first falling and then rising is identified, the fluctuation segments exceeding the set amplitude threshold are screened, and the return window list is constructed. S003, within the constraints of the foldback window list, performs edge fidelity filtering. By dynamically setting the filtering window range, it suppresses trend smoothing across foldback segments, preserves the waveform slope of the rising segment, eliminates peak clipping effects, and outputs enhanced trend data. S004 After completing the edge fidelity filtering process, the temperature sensing data is frozen to decouple the change paths of fast and slow variables. The heat accumulation value is generated by the current square time accumulation calculation method, and the heat protection threshold is constructed to improve the overload risk level. S005, after the overload risk level is increased, the weighting coefficients of the multi-source sensing channels are reallocated, the weight of current and power data is increased in the foldback window, the weight of temperature and environmental data is reduced, and a pre-warning index is generated as an alarm triggering lead signal. S006 After obtaining the early warning index, the residual changes between the trend prediction data and the actual measurement data are compared, the threshold parameters, sliding window width and sensing channel weighting coefficients are dynamically updated, and the updated parameter set is written into the foldback sensitive baseline to complete the closed-loop risk alarm process from perception, identification, processing to control. Step S005 includes: Based on the risk level corresponding to the heat accumulation value within the return window, the weighting coefficient of the sensing channel is dynamically adjusted, increasing the weight of the current channel and power channel to no less than 1.5, and decreasing the weight of the conductor temperature channel and ambient temperature channel to no more than 0.

6. The current, power, conductor temperature and ambient temperature data at each sampling moment within the foldback window are multiplied by their corresponding adjusted weights, and the weighted values ​​are summed to generate a continuous weighted risk data sequence. The normalization interval of each sensing channel is determined by using historical data within 24 hours. The weighted risk data sequence is standardized to generate a forward warning index distributed in the range of 0 to 100. The early warning index is compared with the set risk classification threshold. If the index exceeds the preset threshold for three consecutive sampling periods, the current return window is marked as an abnormal and continuously developing state, and the weight configuration and index evolution trajectory are recorded.

2. The method for overload risk alarm of substation based on multi-source sensing data fusion according to claim 1, characterized in that, Step S001 includes: The preset current sensor, voltage sensor, load detection device, conductor temperature sensor, equipment operation status acquisition device, and environmental temperature and humidity sensing device are capable of receiving external micro-amplitude synchronization pulses and recording timestamps; In the initial stage of data acquisition, the central control unit sends a synchronization pulse signal with a voltage amplitude of less than 100 millivolts and a duration of less than 10 milliseconds to each sensing channel, and collects the receiving timestamps fed back by each channel and calculates the time difference. By statistically analyzing the sampling clock offset pattern of each channel within multiple consecutive synchronization pulse cycles, a trend of time difference change per second is formed, and a second-level alignment vector is generated for unified time base correction. The timestamps of all channel sampled data are adjusted based on the second-level alignment vector. The data time axis is corrected by time shifting or interpolation. A foldback sensitive baseline based on a unified standard time is constructed to provide time consistency assurance for subsequent abnormal fluctuation identification.

3. The method for overload risk alarm of substation based on multi-source sensing data fusion according to claim 1, characterized in that, Step S002 includes: Based on the time-aligned sensing data of the current channel, load power channel and conductor temperature channel, a sliding window is divided by sampling once per second, with a window width of 30 seconds and a step size of 1 second, and the continuous derivative sequence within each window is calculated. Identify whether there is a fluctuation pattern formed by a combination of continuous falling segments and continuous rising segments in each window. Each segment change should have no less than 3 sampling points, and the amplitude should exceed 8% of the rated current value. The time interval between the preceding and following segments should be no less than 5 seconds. For windows that initially meet the conditions, the load power channel and conductor temperature channel are retrieved and a consistency verification is performed within the same time interval to confirm that the power fluctuation direction is consistent and the amplitude exceeds 10% of the rated value, and the temperature channel rises by more than 1.5 degrees Celsius within 10 seconds after the end of the window. The time periods that meet the multi-channel verification conditions are compiled into a structured list of return windows, and the start and end times of the windows, channel numbers, change ranges and verification results are recorded as input for subsequent trend retention and risk identification.

4. The method for overload risk alarm of substation based on multi-source sensing data fusion according to claim 1, characterized in that, Step S003 includes: Extract the start time, end time, channel type, and fluctuation characteristics of each return window, and extract the corresponding time period from the original sensing data and extend it by 1 second before and after as a protection area; The sliding window width is dynamically set based on the maximum value of the derivative of the turnaround segment and the time span, and continuous filtering is performed using a weighted average method, with the weights gradually decreasing from the center of the window to both ends. For continuous sampling points in the upward trend segment, a slope protection mechanism is enabled. When three sets of consecutive positive difference values ​​are met and the amplitude is not lower than the rated threshold, filtering is paused and the original data is retained. Data that is detected as peak points are not modified. After all the turnaround segments are processed, the filtering results are embedded into the original data sequence, and linear interpolation is performed with a 2-second buffer before and after to ensure data continuity.

5. The method for overload risk alarm of substation based on multi-source sensing data fusion according to claim 1, characterized in that, Step S004 includes: Extract the sensing data sequence after edge fidelity filtering, acquire conductor temperature, equipment casing temperature and ambient temperature data within the corresponding time period of the foldback window, and keep the most recent effective temperature value before the start of the window unchanged within the window; The current sensing data sampled per second within the foldback window is squared, and all squared current values ​​are accumulated to calculate the current square time product as the heat input. By combining the equipment model, cable parameters and ventilation conditions to match the equipment's heat capacity carrying capacity, the product of the square of the current over time is compared with the heat capacity threshold to classify the overload risk levels as Level 1, Level 2 and Level 3. The risk level, start and end time, peak current and freezing temperature corresponding to each turnaround window are written into the turnaround window list for use in subsequent dynamic weight adjustment and early warning index generation steps.

6. The method for overload risk alarm of substation based on multi-source sensing data fusion according to claim 1, characterized in that, Step S006 includes: Calculate the residual per second between the trend prediction data and the actual perceived data within the loop window, generate a residual change sequence, and assess the degree of deviation of the prediction model; The alarm parameters are adjusted based on the residual change results, including lowering the over-alarm thresholds for current and power, compressing the sliding window width, increasing the weight of fast variable channels and simultaneously decreasing the weight of slow variable channels. Each parameter adjustment result is recorded as an independent parameter set, including the return window number, residual statistics, weights before and after the update, and changes in window width and alarm threshold; The independent parameter set is written into the return sensitive baseline by timestamp and used as the priority configuration basis for the next round of risk assessment and perception preprocessing operations to achieve closed-loop adaptive control of the alarm process.

7. A substation overload risk alarm device based on multi-source sensing data fusion, used to implement the substation overload risk alarm method based on multi-source sensing data fusion as described in any one of claims 1-6, characterized in that, It includes a data alignment module, a foldback identification module, a fidelity filtering module, a variable decoupling and risk modeling module, a weighted adjustment and early warning module, and a closed-loop control module. The data alignment module injects micro-amplitude synchronization pulses during the multi-source sensing data acquisition stage in the operation of the substation, collects pulse feedback results, calculates the acquisition clock offset, constructs a second-level alignment vector, realizes timing synchronization between different sensing channels, and establishes a foldback sensitive baseline. The foldback identification module, under the constraint of the foldback sensitive baseline, calculates the continuous derivative sequence based on the second-level aligned multi-source sensing data using a sliding window method, identifies the fluctuation pattern of first falling and then rising, filters fluctuation segments that exceed the set amplitude threshold, and constructs a foldback window list; The fidelity filtering module performs edge fidelity filtering within the constraints of the foldback window list. By dynamically setting the filtering window range, it suppresses trend smoothing across foldback segments, preserves the waveform slope of the rising segment, eliminates peak clipping effects, and outputs enhanced trend data. The variable decoupling and risk modeling module performs a freezing operation on temperature-related sensing data after completing edge-fidelity filtering, thereby decoupling the change paths of fast and slow variables. It generates heat accumulation values ​​by using the current square time cumulative calculation method, constructs a heat protection threshold, and improves the overload risk level. The weighted adjustment and early warning module reallocates the weighting coefficients of the multi-source sensing channels after the overload risk level is increased. In the foldback window, the weight of current and power data is increased, while the weight of temperature and environmental data is decreased, and a pre-warning index is generated as an alarm triggering lead signal. The closed-loop control module, after obtaining the early warning index, compares the residual changes between the trend prediction data and the actual measurement data, dynamically updates the threshold parameters, sliding window width and sensing channel weighting coefficients, and writes the updated parameter set into the foldback sensitive baseline, thus completing the closed-loop risk alarm process from perception, identification, processing to control.

Images