Deep learning time series anomaly detection system for energy demand forecasting

Abstract

The present application relates to the technical field of deep learning, and discloses a deep learning time sequence anomaly detection system for energy demand prediction, which comprises the following steps: first, taking the tapping action of the on-load voltage regulating switch as a trigger time point to divide the operation interval, and shielding the interval in the training of the basic load prediction network to extract the pure load deviation sequence; then, combining the bus voltage sequence to calculate the transition hysteresis area index, and inputting the index into the conditional variational autoencoder to learn the operation reference space and extract the interval state vector; tracking the interval state vector through the gated recurrent network to establish the equipment state history sequence, and calculating the threshold-free anomaly continuous evaluation value by using the historical distribution probability; finally, the system performs soft cleaning on the interval data according to the evaluation value, and outputs the final demand prediction result. The present application effectively solves the problem of prediction baseline drift caused by the distortion of the voltage regulating device, and improves the robustness of anomaly detection.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a deep learning-based time series anomaly detection system for energy demand forecasting. Background Technology

[0002] With the popularization of renewable energy and the acceleration of urban electrification, especially the large-scale integration of new electrical equipment such as electric vehicles, power loads are exhibiting strong nonlinearity, non-stationarity, and multi-frequency characteristics. This makes load forecasting and timing anomaly detection core tasks for modern distribution network operation and dispatching. In current distribution networks, on-load tap changers are widely used to regulate bus voltage to cope with fluctuations on both the source and load sides. However, recent research shows that during long-term operation, the mechanical components of such tap changers, such as contacts and switching devices, inevitably experience wear or degradation.

[0003] At this point, the mixed loads of constant impedance, constant current, and constant power widely present on the downstream user side are not entirely insensitive to voltage. When the voltage regulating equipment experiences slight oscillations or tailing during the switching transition process due to component wear, the bus voltage path changes accordingly, and the instantaneous metered power of the downstream load also shifts accordingly. This shift does not originate from changes in the actual electricity consumption habits of users, but rather from pseudo-demands distortedly written into the timing data by upstream equipment.

[0004] Existing deep learning-based demand forecasting systems have significant shortcomings in handling the above situations: when extracting time-series patterns, the forecasting model is highly susceptible to learning these pseudo-demand offsets written by the equipment as normal load fluctuation patterns, leading to a severe drift in the baseline of demand forecasting. Baseline drift not only compromises the accuracy of energy forecasting but also triggers misjudgments by time-series anomaly detection systems, causing genuine equipment malfunctions to be masked or incorrectly attributed to ordinary load fluctuations. Summary of the Invention

[0005] This invention provides a deep learning-based time series anomaly detection system for energy demand forecasting, which solves the technical problems mentioned in the background art.

[0006] This invention provides a deep learning-based time series anomaly detection system for energy demand forecasting, comprising: The operating interval division module is used to divide the operating interval by using the tap-out action of the on-load tap changer as the trigger point and the midpoint of the adjacent tap-out actions, and to obtain tap-out interval data containing the bus voltage sequence. The basic load forecasting module is used to construct the basic load forecasting network. During training, the data of the tap interval is masked to output a load deviation sequence that removes the influence of voltage regulation. The hysteresis index calculation module is used to calculate the transition hysteresis area index based on the bus voltage sequence and the load deviation sequence within the tap interval data. The operation mode learning module is used to input the tap interval data and the transition hysteresis area index into the conditional variational autoencoder, learn the operation reference space of the tap interval data and extract the interval state vector. The device status tracking module is used to input the interval status vector in chronological order to establish a historical sequence of device status. The anomaly quantification assessment module is used to synthesize a comprehensive deviation based on the deviation of the interval state vector and the deviation of the transition hysteresis area index, and to calculate the anomaly continuous assessment value through historical distribution probability. The demand forecasting correction module is used to smooth and correct the tap interval data based on the abnormal continuous evaluation value to obtain the corrected load sequence, and input the corrected load sequence into the basic load forecasting network to output the final demand forecasting result.

[0007] The operating interval division module is used to divide the operating interval by using the tap-out action of the on-load tap changer as the trigger point and the midpoint of the adjacent tap-out actions, and to obtain tap-out interval data containing the bus voltage sequence. The basic load forecasting module is used to construct the basic load forecasting network. During training, the data of the tap interval is masked to output a load deviation sequence that removes the influence of voltage regulation. The hysteresis index calculation module is used to calculate the transition hysteresis area index based on the bus voltage sequence and the load deviation sequence within the tap interval data. The operation mode learning module is used to input the tap interval data and the transition hysteresis area index into the conditional variational autoencoder, learn the operation reference space of the tap interval data and extract the interval state vector. The device status tracking module is used to input the interval status vector in chronological order to establish a historical sequence of device status. The anomaly quantification assessment module is used to synthesize a comprehensive deviation based on the deviation of the interval state vector and the deviation of the transition hysteresis area index, and to calculate the anomaly continuous assessment value through historical distribution probability. The demand forecasting correction module is used to smooth and correct the tap interval data based on the abnormal continuous evaluation value to obtain the corrected load sequence, and input the corrected load sequence into the basic load forecasting network to output the final demand forecasting result.

[0008] The beneficial effects of this invention are as follows: by using the tap changer action of the on-load tap changer as the clock anchor point, extracting the transition hysteresis area index, and using a deep generation model and cyclic network to track the chronic degradation trajectory of the equipment, it is possible to effectively remove pseudo-demand anomalies written into the load timing due to equipment switching distortion; this invention gets rid of the limitation of manually preset thresholds, realizes continuous evaluation of equipment status and smooth correction of the demand forecast benchmark sequence, which not only greatly improves the timing prediction accuracy of the distribution network under complex load conditions, but also realizes early and accurate identification of potential degradation hazards of tap changer equipment. Attached Figure Description

[0009] Figure 1 This is a block diagram of the deep learning time series anomaly detection system for energy demand forecasting of the present invention. Detailed Implementation

[0010] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0011] like Figure 1 As shown, a deep learning-based time series anomaly detection system for energy demand forecasting includes: The operating interval division module is used to divide the operating interval by using the tap-out action of the on-load tap changer as the trigger point and the midpoint of the adjacent tap-out actions, and to obtain tap-out interval data containing the bus voltage sequence. The basic load forecasting module is used to construct the basic load forecasting network. During training, the data of the tap interval is masked to output a load deviation sequence that removes the influence of voltage regulation. The hysteresis index calculation module is used to calculate the transition hysteresis area index based on the bus voltage sequence and the load deviation sequence within the tap interval data. The operation mode learning module is used to input the tap interval data and the transition hysteresis area index into the conditional variational autoencoder, learn the operation reference space of the tap interval data and extract the interval state vector. The device status tracking module is used to input the interval status vector in chronological order to establish a historical sequence of device status. The anomaly quantification assessment module is used to synthesize a comprehensive deviation based on the deviation of the interval state vector and the deviation of the transition hysteresis area index, and to calculate the anomaly continuous assessment value through historical distribution probability. The demand forecasting correction module is used to smooth and correct the tap interval data based on the abnormal continuous evaluation value to obtain the corrected load sequence, and input the corrected load sequence into the basic load forecasting network to output the final demand forecasting result.

[0012] Preferably, the tap changer action is used as the trigger point, and the operating interval is divided using the midpoint of adjacent tap changes to obtain tap interval data containing the bus voltage sequence, including: Obtain the tap position sequence, and determine the moment when the value changes in the tap position sequence as the trigger point. The calculation formula is as follows: in, For the first The triggering time point mentioned below For time moments, The value of the tap position sequence corresponding to the stated time. The value of the tap position sequence corresponding to the previous time point; The median time between the previous triggering time and the current triggering time is rounded down to serve as the starting boundary of the running interval. Similarly, the median time between the current triggering time and the next triggering time is rounded down to serve as the ending boundary of the running interval. The calculation formula is as follows: in, The starting boundary of the operating range is the interval. The terminus of the operating range is the terminus of the range termination boundary. This represents the floor function. For the first The triggering time point mentioned below For the first The triggering time point described below; The active power sequence, the bus voltage sequence, and the external environment feature sequence between the start boundary and the end boundary of the interval are extracted to jointly construct the tap interval data. The calculation formula is as follows: in, For the first The aforementioned tap interval data The active power sequence at the specified time point. The bus voltage sequence at the specified time point. The external environment feature sequence at the specified time point.

[0013] The tap position sequence is a sequence of on-load tap changers recorded in a uniform sampling order. It can be obtained through an on-load tap changer position feedback device.

[0014] The trigger point is the time position where the numerical value of the tap sequence changes.

[0015] The time point is a unified time index for all sampled data.

[0016] The value of the tap position sequence corresponding to the previous time point is the tap position value of the sampling time point before the current time point.

[0017] The previous trigger point is the time position of the most recent tapping action before the current trigger point, used to determine the starting boundary of the interval.

[0018] The next trigger point is the time position of the most recent tapping action after the current trigger point, used to determine the interval termination boundary.

[0019] The median time value is the intermediate time value between two adjacent trigger times, used to distribute the impact of a tapping action across the two running intervals on the time axis.

[0020] The starting boundary of the interval is the left boundary position of the current running interval, which is used to determine the starting sample of the data in the current sub-interval.

[0021] The interval termination boundary is the right boundary position of the current running interval, used to determine the termination sample of the current tap interval data.

[0022] The operating interval is a time segment established around a single tap change operation, used to carry the complete process of bus voltage changes and load response before and after the tap change operation.

[0023] The active power sequence is a sequence of load active power changes recorded according to a unified time base. It can be obtained through feeder metering devices, transformer metering devices, or distribution terminals. After collection, it is uniformly divided by the rated power base value of the distribution network to perform dimensionless processing, transforming it into a pure numerical variable, thereby meeting the dimensional requirements when performing natural logarithmic calculations on the variance term in the subsequent basic network loss function.

[0024] A bus voltage sequence is a sequence of bus voltage changes recorded according to a uniform time base. It can be obtained using a bus voltage measuring device or a voltage transformer acquisition device.

[0025] The external environmental characteristic sequence is a sequence of external influencing factors related to load changes. It can be obtained through environmental monitoring devices, meteorological information sources, and calendar information sources.

[0026] Tap interval data is a set of event data consisting of active power sequence, bus voltage sequence, and external environment characteristic sequence between the start boundary and end boundary of the interval.

[0027] The sequence number is used to indicate the order of the trigger point, the operating interval, and the tap interval data.

[0028] In detail, because the tap-changing action of the on-load tap changer is a discrete equipment action, a fixed time window cannot guarantee that the bus voltage sequence and active power sequence before and after the same tap-changing action fall into the same operating range. Therefore, using the tap-changing action of the on-load tap changer as the trigger point can make the operating range revolve around the actual equipment behavior.

[0029] In detail, because the impact of a single tapping action on the load includes both the transition preparation phase before the action and the voltage recovery phase after the action, using only the trigger point itself as the boundary would result in an excessively narrow interval or overlapping intervals. However, by using the midpoint of adjacent tapping actions to divide the operating interval, the impact of adjacent tapping actions can be evenly separated without introducing additional manual thresholds. Therefore, this division method is not a normal time slice, but an interval construction method closely related to equipment events.

[0030] In detail, since the tap position sequence is a direct discrete representation of the working state of the on-load tap changer, as long as a value changes in the tap position sequence, it indicates that a new tapping action has entered the current sampling time. Therefore, determining the time when the value changes in the tap position sequence as the trigger time point can convert the mechanical action information into a recognizable time anchor point.

[0031] In detail, because the current operating interval needs to avoid both the tail effect of the previous tapping action and the leading effect of the next tapping action, the median time value between the previous triggering time point and the current triggering time point is used as the starting boundary of the interval, and the median time value between the current triggering time point and the next triggering time point is used as the ending boundary of the interval. This ensures that the current operating interval remains centered on the current triggering time point on the time axis.

[0032] In detail, the tap position sequence is obtained as follows: The tap position value corresponding to each sampling moment is read from the tap position feedback loop of the on-load tap changer, and this value is written into the timing data with a unified timestamp. When the tap position feedback is a discrete contact state, the contact state is first mapped to a unique tap position value, and then the tap position sequence is formed according to the sampling order.

[0033] In detail, the starting and ending boundaries of the first and last running intervals are determined as follows: When the current running interval corresponds to the first trigger time, the data start time is taken as the starting boundary of the interval. When the current running interval corresponds to the last trigger time, the data end time is taken as the ending boundary of the interval. When there are available trigger time points on both sides of the first or last trigger time point, the boundary is still determined by the median value of the time between adjacent trigger time points.

[0034] In detail, the calculation method for the median time value is as follows: First, the previous, current, and next trigger points are uniformly represented as time index values ​​under the same sampling benchmark. Then, the index values ​​of two adjacent trigger points are averaged. Finally, the average result is rounded down to obtain the interval start boundary and interval end boundary corresponding to each sampling point. When running online and the next trigger point has not yet occurred, the preset maximum observation duration is used to replace the median time value to determine the interval end boundary. To avoid the shielding range from expanding infinitely due to excessively long intervals between adjacent actions, when the calculated interval time span is greater than the maximum duration of voltage regulation physical recovery, the interval end boundary is truncated according to this maximum duration, so that the training shielding coefficient of the no-action steady-state time period remains at a value of one.

[0035] In detail, the external environment feature sequence is composed as follows: temperature, humidity, wind speed, rainfall, time period, weekday, holidays, and external operating status information related to load changes are combined in a unified time sequence to form a multi-dimensional external environment feature sequence, and each time point corresponds to a unique set of external environment feature values.

[0036] In detail, the time synchronization method for active power sequence, bus voltage sequence, and external environmental characteristic sequence is as follows: First, a unified sampling period is selected as a common time reference. Then, data from different sampling periods are resampled onto this common time reference. For data with a sampling frequency higher than the common time reference, a single value is obtained by aggregating samples within the common time reference. For data with a sampling frequency lower than the common time reference, the most recent valid time is retained until the next update time. Then, the data is aligned step by step according to the same time index, and the extracted sequence is resampled to a preset unified standard sample number through time-step linear interpolation. This eliminates the scale distortion effect caused by the duration difference of different operating intervals on subsequent interval normalization operations, ultimately forming the tap interval data.

[0037] Preferably, a basic load forecasting network is constructed, and the tap interval data is masked during training to output a load deviation sequence that removes the influence of voltage regulation, including: The historical active power sequence and the external environment feature sequence from history to the present are input into a temporal convolutional network to extract temporal context features. The calculation formula is as follows: in, For the aforementioned temporal context features, The mapping function of the temporal convolutional network is... The active power sequence is the historical sequence from time one to the previous time. The sequence of external environmental features from time one to the present; The temporal context features and the previous time-series recurrent hidden state vector are input into the gated recurrent unit to extract the current recurrent hidden state vector. The calculation formula is as follows: in, Let be the current cyclic hidden state vector. The mapping function for the gated loop unit. Let be the cyclic hidden state vector of the previous time step; The current cyclic hidden state vector is linearly mapped to obtain the baseline mean and baseline variance of demand. The calculation formula is as follows: in, The benchmark mean of the demand is... The required baseline variance, For linear mapping weights, For linear mapping bias; Between the start boundary and the end boundary of the interval, the training masking coefficient is set to zero; in other time ranges, the training masking coefficient is set to one. The calculation formula is as follows: in, The training masking coefficient is... For the current moment, This is the starting boundary of the interval. This is the termination boundary of the interval; Using the training masking coefficient, loss calculations are performed on the current active power sequence, the baseline demand mean, and the baseline demand variance to obtain the basic network loss value for updating network parameters. The calculation formula is as follows: in, The basic network loss value, The current active power sequence; The load deviation sequence is obtained by subtracting the current active power sequence from the baseline demand mean, and the calculation formula is as follows: in, This refers to the load deviation sequence.

[0038] The base load forecasting network is a forecasting model used to learn a normal demand baseline that does not contain the effects of voltage regulation writes.

[0039] Temporal convolutional networks are networks that extract temporal features from historical samples. They are used to extract temporal context features from historical active power sequences and historical to current external environment feature sequences.

[0040] Temporal context features are temporal dependency representations output by temporal convolutional networks, used to summarize load evolution and changes in the external environment prior to the current moment.

[0041] A gated loop unit is a temporal modeling unit that uses a gating mechanism to update the loop hidden state vector, combining the current context with historical memory.

[0042] The cyclic hidden state vector of the previous time step is the internal state representation retained from the previous time step, used to pass on historical memory to the current time step.

[0043] The current cyclic hidden state vector is the internal state representation output by the gated cyclic unit at the current moment, which is used to carry the currently fused temporal memory information.

[0044] The baseline mean demand is the central estimate of the normal demand level at the current moment by the base load forecasting network.

[0045] Demand baseline variance is a discrete estimate by the base load forecasting network of the degree of normal demand fluctuation at the current moment.

[0046] The natural logarithm of the required baseline variance is the value obtained by logarithmically transforming the required baseline variance. It is used to make the training process more stable and to ensure that the required baseline variance is positive.

[0047] Linear mapping weights are weight parameters that map the current cyclic latent state vector to the demand baseline mean and demand baseline variance.

[0048] Linear mapping bias is a bias parameter superimposed on the mapping result during the linear mapping process.

[0049] The training masking coefficient controls whether the current sample participates in the calculation of the basic network loss value. Preferably, it is 0 between the start and end boundaries of the interval, and 1 in other time ranges. This is because samples within the operating interval contain voltage regulation effects and must be masked out from the normal pattern learning of the basic load prediction network.

[0050] The base network loss value is a training target value used to measure the degree of difference between the output of the base load prediction network and the current active power sequence.

[0051] Network parameters are the set of all adjustable parameters that need to be updated during the training of the basic load prediction network.

[0052] The load deviation sequence is the deviation sequence obtained by subtracting the current active power sequence from the demand baseline mean, and is used to represent the remaining change after removing the normal demand baseline.

[0053] In detail, because the data packets in the tap interval contain voltage writing effects caused by the tap changer action, if the data in this interval is directly used as normal samples for training, the basic load forecasting network will learn the pseudo-demand written by the equipment as a normal demand pattern. Therefore, when constructing the basic load forecasting network, the data in the tap interval must be masked to ensure that the output load deviation sequence mainly reflects the effects of voltage regulation and abnormalities.

[0054] In detail, because the historical active power sequence contains both slow-changing trends and periodic changes, in order to prevent abnormal data written by the device from polluting the benchmark model through autoregressive feedback, the load sequence smoothed and corrected by historical steps is used as the historical active power sequence here before inputting it into the network. The external environment feature sequence from history to the present contains external driving information such as weather and time period. Therefore, inputting both into the temporal convolutional network can extract the available temporal context features at the current moment without using future samples.

[0055] In detail, since the temporal context features output by the temporal convolutional network alone cannot continuously preserve long-term memories, the temporal context features and the recurrent hidden state vector of the previous time step are input into the gated recurrent unit. This allows the memory of the previous time step to be jointly encoded with the new information of the current time step. Therefore, the current recurrent hidden state vector can simultaneously reflect local changes and historical dependencies.

[0056] In detail, because demand forecasting not only needs to estimate the normal demand level, but also the fluctuation range around that level, the current cyclic hidden state vector is linearly mapped to obtain the demand baseline mean and demand baseline variance, which allows the base load forecasting network to output both the central value and the uncertainty.

[0057] In detail, because the training shielding coefficient is 0 within the operating interval and 1 in other time ranges, when using the training shielding coefficient to calculate the loss of the current active power sequence, the mean of the demand benchmark, and the variance of the demand benchmark, the samples within the operating interval will not push the network parameters to update in the direction of false demand. Therefore, the basic network loss value has the function of eliminating the pollution of voltage regulation events.

[0058] In detail, the network structure of the temporal convolutional network is as follows: the temporal convolutional network is set up as a convolutional extraction structure that processes historical samples layer by layer along the time axis. The input end receives the historical active power sequence and the external environment feature sequence from the past to the present. The output end provides the temporal context features aligned with the current time and ensures that the output at any time only uses the current and previous sample information.

[0059] In detail, the initialization method of the gated cyclic unit is as follows: at the first valid time step, the cyclic hidden state vector of the previous time step is set to a vector of all zeros, and then updated step by step from the vector of all zeros. After that, the cyclic hidden state vector output by the previous time step is used as the input of the current time step at each time step.

[0060] In detail, the accumulation method of the basic network loss value is as follows: calculate the loss term corresponding to each individual sample in chronological order for all training samples, then use the training masking coefficient to multiply and filter each loss term, and finally sum or average all the filtered loss terms to obtain the basic network loss value.

[0061] In detail, the network parameters are updated as follows: all network parameters in the basic load forecasting network are adjusted in reverse according to the basic network loss value, so that the demand baseline mean gradually approaches the normal demand level that is not affected by voltage regulation, and the demand baseline variance gradually reflects the normal demand fluctuation range.

[0062] In detail, the training masking coefficient is applied as follows: First, mark whether the current time is in the running interval according to the start boundary and end boundary of the interval. Then, assign the training masking coefficient corresponding to the time in the running interval to 0, and assign the training masking coefficient corresponding to the time not in the running interval to 1. Finally, multiply the training masking coefficient directly onto the loss term corresponding to each time.

[0063] Preferably, the transition hysteresis area index is calculated based on the bus voltage sequence and the load deviation sequence within the tap interval data, including: Subtracting the load deviation value corresponding to the starting boundary of the interval from the load deviation sequence yields the centered load deviation sequence. Similarly, subtracting the bus voltage value corresponding to the starting boundary of the interval from the bus voltage sequence yields the centered bus voltage sequence. The calculation formula is as follows: in, For the centered load deviation sequence, The centralized bus voltage sequence, The load deviation sequence, The bus voltage sequence, The load deviation value corresponding to the starting boundary of the interval. The bus voltage value corresponding to the starting boundary of the interval; The normalized load deviation sequence is obtained by dividing the centered load deviation sequence by the sum of the square root of the sum of squares between the start boundary and the end boundary of the interval and the numerical stability constant. Similarly, the normalized bus voltage sequence is obtained by dividing the centered bus voltage sequence by the sum of the square root of the sum of squares between the start boundary and the end boundary of the interval and the numerical stability constant. The calculation formula is as follows: in, For the normalized load deviation sequence, The normalized bus voltage sequence, This is the starting boundary of the interval. This is the termination boundary of the interval. The numerical stability constant is denoted as . The normalized load deviation sequence and the normalized bus voltage sequence corresponding to the starting boundary of the interval are appended to the end of the sequence to form a closed endpoint. The arc length of the closed curve is calculated based on the adjacent differences between the normalized load deviation sequence and the adjacent differences between the normalized bus voltage sequence. The calculation formula is as follows: in, Let the arc length of the closed curve be . The normalized bus voltage value that forms the closed endpoint at the end of the sequence. The normalized load deviation value is the value at which the closed endpoint is formed at the end of the sequence; Based on the arc length of the closed curve, the adjacent mean values ​​of the normalized load deviation sequence, and the adjacent differences of the normalized bus voltage sequence, the transition hysteresis area index is calculated using the discrete integral area formula. The calculation formula is as follows: in, The transition hysteresis area index is mentioned above.

[0064] The load deviation value corresponding to the starting boundary of the interval is the load deviation sequence value at the left boundary of the operating interval.

[0065] The bus voltage value corresponding to the starting boundary of the interval is the bus voltage sequence value at the left boundary of the operating interval.

[0066] The centered load deviation sequence is a sequence obtained by subtracting the load deviation value corresponding to the starting boundary of the interval from the load deviation sequence, and is used to eliminate absolute level differences.

[0067] The centralized bus voltage sequence is a sequence obtained by subtracting the bus voltage value corresponding to the starting boundary of the interval from the bus voltage sequence, and is used to eliminate differences in absolute voltage levels.

[0068] The numerical stability constant is an extremely small positive number that prevents the denominator from becoming zero during normalization and bias correction calculations. A value of 0.000001 is preferred. This value is small enough that it will not substantially change the normalization result or the overall bias ranking, but it can prevent denominator instability.

[0069] The normalized load deviation sequence is a sequence obtained by normalizing the centered load deviation sequence according to the interval energy, and is used to eliminate the difference in amplitude scale between different operating intervals.

[0070] The normalized bus voltage sequence is a sequence obtained by normalizing the centered bus voltage sequence according to the interval energy, and is used to eliminate the voltage step difference between different operating intervals.

[0071] The closed endpoint is the end point of the closed path formed by appending the normalized load deviation sequence and the normalized bus voltage sequence corresponding to the starting boundary of the interval to the end of the sequence.

[0072] The normalized load deviation value that forms a closed endpoint at the end of the sequence is the endpoint value of the closed endpoint in the load deviation dimension.

[0073] The normalized bus voltage value that forms a closed endpoint at the end of the sequence is the endpoint value of the closed endpoint in the bus voltage dimension.

[0074] The arc length of a closed curve is the total path length of the normalized bus voltage sequence and the normalized load deviation sequence on a two-dimensional trajectory after closure.

[0075] The adjacent difference is the numerical difference result of two adjacent sampling points on the same normalized sequence.

[0076] The adjacent mean is the average result of two adjacent sampling points on the same normalized sequence.

[0077] The discrete integral area formula is a calculation method that uses adjacent mean values ​​and adjacent differences to discretely solve for the directional area of ​​a closed curve.

[0078] The transition hysteresis area index is an indicator that characterizes whether the bus voltage sequence and the load deviation sequence form a hysteresis loop and the strength of it within the operating range. In its calculation process, it is normalized by dividing by the arc length of the closed curve, and the two-dimensional area is transformed into the average hysteresis width per unit trajectory length, thus eliminating the interference of severe path oscillations on the assessment of hysteresis characteristics.

[0079] In detail, because the load deviation sequence and bus voltage sequence in different operating ranges may have different absolute starting levels, directly comparing the original values ​​would mistake the static bias for a dynamic trajectory difference. Therefore, subtracting the corresponding values ​​of the starting boundary of the range from the load deviation sequence and the bus voltage sequence respectively can focus the comparison on the relative change process within the operating range.

[0080] In detail, because the voltage step size and load deviation amplitude may be significantly different in different operating ranges, if amplitude normalization is not performed, the transition hysteresis area index will be dominated by large value events. Therefore, dividing the centered load deviation sequence and bus voltage sequence by the range energy and adding the numerical stability constant can bring different operating ranges to a comparable scale. Thus, the normalized geometric area mainly reflects the trajectory shape rather than the original amplitude size.

[0081] In detail, since the directed area can only be established on a closed path, and the original trajectory within the operating interval only gives an open path from the starting boundary of the interval to the ending boundary of the interval, the normalized load deviation sequence and the normalized bus voltage sequence corresponding to the starting boundary of the interval are appended to the end of the sequence to form a closed endpoint, which can transform the time-series response trajectory into a closed curve with a calculable directed area.

[0082] In detail, because the directional area alone is affected by the path length, and the number of sampling points and path shape are not exactly the same in different operating intervals, the transition hysteresis area index can be calculated by using the discrete integral area formula based on the arc length of the closed curve, the adjacent mean of the normalized load deviation sequence, and the adjacent difference of the normalized bus voltage sequence. This can further normalize the hysteresis characteristics under the unit path scale, so the index is more suitable for cross-interval comparison.

[0083] In detail, the numerical stability constant is determined as follows: It is set to a very small positive number that is at least three orders of magnitude smaller than the typical energy of the normalized sequence, and remains consistent throughout the entire operating range. A value of 0.000001 is preferred, ensuring that it only provides stability when the denominator is close to zero, without altering the relative ordering in the normal range.

[0084] In detail, the method for constructing the closed endpoint is as follows: after the sample corresponding to the termination boundary of the interval, add a new sample, and directly set the normalized load deviation value and normalized bus voltage value of the new sample to the normalized load deviation value and normalized bus voltage value corresponding to the starting boundary of the interval, thereby completing the path closure from the end point back to the starting point.

[0085] In detail, the handling method when the arc length of the closed curve is zero or close to zero is as follows: first calculate the arc length of the closed curve according to the adjacent difference. When the arc length of the closed curve is less than the numerical stability constant, directly replace the arc length of the closed curve with the numerical stability constant, and then continue to calculate the transition hysteresis area index to avoid numerical amplification caused by division by zero or extremely small denominators.

[0086] In detail, the calculation method of the discrete integral area formula is as follows: strictly follow the time sequence from the starting boundary of the interval to the ending boundary of the interval and back to the starting boundary of the interval, accumulate point by point, first calculate the adjacent mean of the normalized load deviation sequence, then multiply it by the adjacent difference of the normalized bus voltage sequence, and finally sum all discrete terms and divide by the arc length of the closed curve, thereby retaining the directional information of the transition hysteresis area index.

[0087] Preferably, the tap interval data and the transition hysteresis area index are input into a conditional variational autoencoder to learn the operating reference space of the tap interval data and extract the interval state vector, including: The normalized bus voltage sequence and the normalized load deviation sequence between the start boundary and the end boundary of the interval are concatenated with the transition hysteresis area index broadcast according to the time length to obtain the comprehensive input tensor. The calculation formula is as follows: in, For the comprehensive input tensor, This is the function for concatenation operations. The normalized bus voltage sequence is the sequence of voltages from the start boundary of the interval to the end boundary of the interval. The normalized load deviation sequence is the sequence of load deviations between the starting boundary and the ending boundary of the interval. The transition hysteresis area index is mentioned above. It is a one-way broadcast vector; The integrated input tensor is input into the encoding network of the conditional variational autoencoder, and the state mean parameter and state variance parameter are extracted. The calculation formula is as follows: in, The state mean parameter, The state variance parameter is... The standard deviation of the state. This is the mapping function of the encoding network; Based on the state mean parameter, the state variance parameter, and the preset standard normal distribution random parameter, reparameterized sampling is performed to extract the interval state vector. The calculation formula is as follows: in, Let be the interval state vector. The symbol for element-wise multiplication. Let be the random parameter of the standard normal distribution; The interval state vector and the transition hysteresis area index are input into the decoding network of the conditional variational autoencoder to obtain the reconstructed output tensor, calculated as follows: in, For the reconstructed output tensor, This is the mapping function of the decoding network; Based on the mean square error between the integrated input tensor and the reconstructed output tensor, and the relative entropy between the encoding distribution and the standard normal distribution, the spatial learning loss is calculated to update the network parameters and learn the operating baseline space. The calculation formula is as follows: in, Let the spatial learning loss value be... The mean square error between the synthesized input tensor and the reconstructed output tensor is denoted as . The function for calculating the relative entropy, For the coding distribution, It is the standard normal distribution.

[0088] The conditional variational autoencoder is a generative model that learns the normal distribution of data in the tap interval by using the transition hysteresis area index as conditional information.

[0089] The transition hysteresis area index broadcast by time length is a conditional sequence formed by copying the transition hysteresis area index corresponding to a single operating interval to each time position within the interval.

[0090] The all-one broadcast vector is a constant vector used to extend the transition hysteresis area index to the entire length of the running interval. Preferably, it is a vector with a length equal to the number of samples between the start and end boundaries of the interval, and all elements are 1. This vector can repeat a single index value to each time position of the synthesized input tensor time-by-time.

[0091] The composite input tensor is an event input representation obtained by splicing together the normalized bus voltage sequence, the normalized load deviation sequence, and the transition hysteresis area index broadcast according to the time length.

[0092] The concatenation operation function is an operation function that connects multiple inputs of the same time length according to feature dimensions to form a comprehensive input tensor.

[0093] Encoding networks are networks that compress and map the comprehensive input tensor into state mean parameters and state variance parameters.

[0094] The state mean parameter is an estimate of the center position of the interval state vector in the potential space.

[0095] The state variance parameter is an estimate of the degree of dispersion of the interval state vector in the potential space.

[0096] State standard deviation is the standard deviation value corresponding to the state variance parameter, and is used for reparameterized sampling.

[0097] The pre-defined standard normal distribution is the prior distribution that the interval state vector is to approximate in the potential space. Preferably, it is a standard normal distribution with a mean of 0 and a variance of 1. This distribution is simple, stable, and facilitates mapping different operating intervals to a unified operating reference space.

[0098] The standard normal distribution random parameter is a random value drawn from a preset standard normal distribution and is used to achieve reparameterized sampling.

[0099] Reparameterized sampling is a sampling method that generates an interval state vector based on the state mean parameter, state variance parameter, and standard normal distribution random parameter.

[0100] An interval state vector is a latent feature vector that provides a low-dimensional representation of the overall state of an operating interval.

[0101] The decoding network is a network that maps the interval state vector and the transition hysteresis area index back to the reconstructed output tensor.

[0102] The reconstructed output tensor is the tensor output by the conditional variational autoencoder after reconstructing the synthesized input tensor.

[0103] Mean squared error is the squared average of the element-by-element differences between the input tensor and the reconstructed output tensor, and is used to measure the reconstruction quality.

[0104] The encoding distribution is the interval state vector probability distribution obtained by the encoding network under the condition of comprehensive input tensor.

[0105] Relative entropy is a measure of the degree of difference between the encoded distribution and a pre-defined standard normal distribution.

[0106] The spatial learning loss value is the training target value composed of the mean square error and the relative entropy.

[0107] The operating reference space is the latent space representing the event pattern of the normal tap interval, which is learned by the conditional variational autoencoder.

[0108] In detail, since the normalized bus voltage sequence and the normalized load deviation sequence can only reflect the time-by-time trajectory and cannot explicitly express the overall hysteresis intensity of the operating range, the normalized bus voltage sequence and the normalized load deviation sequence are spliced ​​together with the transition hysteresis area index broadcast according to the time length. This allows each time position to simultaneously perceive the local trajectory and the global index. Therefore, the integrated input tensor retains the temporal details while introducing overall conditional information.

[0109] In detail, because what needs to be learned is the typical distribution of tap interval data under normal equipment conditions, rather than simply memorizing the original waveform of a certain fixed interval, the tap interval data and the transition hysteresis area index are input into the conditional variational autoencoder to learn the operating reference space. This can compress different operating intervals into a unified event representation space, so that subsequent comparison of interval state vector deviations has discriminative significance.

[0110] In detail, because if we directly sample randomly from the state mean parameter and the state variance parameter, it will block the continuous transmission of loss to the encoding network during training. Therefore, reparameterized sampling based on the state mean parameter, the state variance parameter and the standard normal distribution random parameter can maintain training differentiability while preserving randomness. Thus, the interval state vector has both probabilistic modeling capability and stable training.

[0111] In detail, because using only the mean square error between the integrated input tensor and the reconstructed output tensor as the target would cause the benchmark space to over-memorize sample details, while using only the relative entropy between the encoding distribution and the standard normal distribution as the target would weaken the event reconstruction capability, the two are combined to form the spatial learning loss value, which can achieve a balance between reconstruction accuracy and potential spatial regularity. Therefore, the benchmark space can not only express the normal event pattern, but also facilitate subsequent continuous evaluation.

[0112] In detail, the tensor dimension organization method of the comprehensive input tensor is as follows: the normalized bus voltage sequence between the start boundary of the interval and the end boundary of the interval is taken as the first feature channel, the normalized load deviation sequence between the start boundary of the interval and the end boundary of the interval is taken as the second feature channel, and the transition hysteresis area index broadcast according to the time length is taken as the third feature channel, and they are stacked in time order to form the comprehensive input tensor.

[0113] In detail, the network structure of the encoding network is as follows: the comprehensive input tensor is input into the encoding network in chronological order. The encoding network first compresses the time changes within the operating interval. During this compression process, a random feature zeroing operation is applied to the data channel corresponding to the transition hysteresis area index broadcast according to the time length to block the direct mapping shortcut from input to reconstruction output. The compression result is then divided into two output branches, one of which corresponds to the state mean parameter and the other of which corresponds to the state variance parameter, to ensure that position estimation and discrete estimation are obtained simultaneously within the same operating interval.

[0114] In detail, the network structure of the decoding network is as follows: the interval state vector and the transition hysteresis area index are input into the decoding network. The decoding network first restores the overall state of the running interval, and then generates a reconstructed output tensor with the same length as the comprehensive input tensor at each time step, so that the reconstructed output tensor and the comprehensive input tensor are consistent in terms of time length and feature arrangement.

[0115] In detail, the method for generating the random parameters of the standard normal distribution is as follows: during training in each running interval, random values ​​are generated element by element from the preset standard normal distribution according to the dimension of the interval state vector. The random values ​​are then multiplied element by element by the state standard deviation and added to the state mean parameter to obtain the interval state vector of the current running interval.

[0116] In detail, the alignment method of the sequence between the start boundary and the end boundary of the variable-length interval is as follows: the time length of the longest running interval in the current batch is used as the uniform length, the shorter running intervals are padded with 0s at the end, and only the real sample positions between the start boundary and the end boundary of the interval are included in the calculation when calculating the mean square error, so as to ensure that different running intervals can enter the conditional variational autoencoder training together.

[0117] Preferably, the interval state vectors are input in chronological order to establish a device state history sequence, including: The device state history sequence from the previous time step is linearly mapped to obtain the state vector for the prediction interval. The calculation formula is as follows: in, Let be the state vector of the prediction interval. This refers to the historical sequence of device states at the previous moment. This is the first weight matrix. This is the first bias vector; The device state history sequence from the previous moment is linearly mapped to obtain the mean and variance of the predicted index. The calculation formula is as follows: in, The mean of the predicted index, To predict the standard deviation of the indicator, Let be the natural logarithm of the variance of the prediction index. This is the second weight matrix. This is the second bias vector; The current interval state vector and the previous device state history sequence are input into the gated recurrent network to update the current device state history sequence. The calculation formula is as follows: in, This is the current historical sequence of the device's state. This is the current interval state vector. The mapping function for the gated recurrent network; Calculate the mean square error of the current interval state vector and the predicted interval state vector; calculate the index probability loss value of the current transition hysteresis area index relative to the mean and variance of the predicted index; and accumulate these two values ​​over multiple consecutive time intervals to obtain the state evolution loss value to update the network parameters. The calculation formula is as follows: in, The state evolution loss value is... This represents the total number of intervals in the time series. Let be the mean square error of the state. The current transition hysteresis area index, The probability loss value of the aforementioned indicator. The squared difference between the current transition hysteresis area index and the mean of the predicted index is... This is the relative deviation value.

[0118] The equipment status history sequence is a time-series representation of the accumulated historical status information of the equipment in the order of its operating interval.

[0119] The predicted interval state vector is the state representation that the current operating interval should have, inferred from the historical sequence of the device state at the previous moment.

[0120] The first weight matrix consists of weight parameters that map the historical sequence of device states at the previous time step to the state vector of the prediction interval.

[0121] The first bias vector is the bias parameter superimposed when the state vector of the prediction interval is obtained by linear mapping.

[0122] The mean of the forecast index is the estimated value of the center of the current transition hysteresis area index, which is predicted based on the historical sequence of equipment status at the previous moment.

[0123] The variance of the prediction index is a discrete estimate of the current transition hysteresis area index, which is predicted based on the historical sequence of the equipment status at the previous moment.

[0124] The standard deviation of a forecast indicator is the standard deviation value corresponding to the variance of the forecast indicator, and is used to characterize the fluctuation scale of the forecast indicator.

[0125] The natural logarithm of the predictor variance is the value obtained by logarithmically transforming the predictor variance, and is used to keep the predictor variance positive and stabilize the training.

[0126] The second weight matrix consists of weight parameters that map the historical sequence of device status at the previous moment to the mean and variance of the prediction index.

[0127] The second bias vector is a bias parameter superimposed when the mean and variance of the predictor are obtained by linear mapping.

[0128] Gated cyclic networks are time-series networks used to update the historical sequence of device states on an interval-by-interval basis.

[0129] The mean square error of the state is the average of the squared difference between the current interval state vector and the predicted interval state vector across all latent feature dimensions. By calculating the dimensional average, it achieves dimensional balance with the dimension-normalized reconstruction error on the same numerical scale.

[0130] The squared difference is the result of squaring the difference between the current transition lag area index and the mean of the predicted index.

[0131] The relative deviation is the normalized deviation obtained by dividing the square of the difference by the variance of the predictor index.

[0132] The probability loss value of the indicator is the loss obtained by adding the natural logarithm of the relative deviation value to the variance of the predicted indicator.

[0133] Multiple consecutive time intervals are a set of operating intervals used to continuously accumulate the mean square error of the state and the probability loss of the index.

[0134] The state evolution loss value is the training target value obtained by summing the mean square error of the state and the index probability loss value over multiple consecutive time intervals.

[0135] The total number of intervals in a time series is the total number of running intervals that have been formed in the current time series.

[0136] In detail, because the interval state vector of a single operating interval can only reflect the local state of that tapping event and cannot reflect the continuous change trend of the equipment state over time, the interval state vector is input in chronological order to establish a historical sequence of equipment states, which can extend the representation of isolated events into a continuous equipment state trajectory.

[0137] In detail, since the current operating interval state and the current transition hysteresis area index are both affected by the states of multiple previous operating intervals, a linear mapping process is performed on the historical sequence of the equipment state at the previous moment to obtain the predicted interval state vector, the mean of the predicted index, and the variance of the predicted index. This allows us to first give the state and index distribution that the current operating interval should have under normal evolution, so the difference between the subsequent actual observation and the prediction has anomaly interpretation significance.

[0138] In detail, because the device state history sequence of the previous moment only contains historical information up to the previous operating interval, while the interval state vector of the current operating interval retains the small offset from the normal prior distribution through the loss weight of relative entropy in the control network, representing new device behavior containing degradation information, the current interval state vector and the device state history sequence of the previous moment are input into the gated recurrent network to update the current device state history sequence. This can include newly occurring tapping events in continuous memory, so the device state history sequence can be updated recursively over time.

[0139] In detail, since equipment anomalies may manifest as the state vector of the interval deviating from the predicted state vector of the interval, or as the transition hysteresis area index deviating from the predicted index mean, the state mean square error and the index probability loss value are accumulated over multiple consecutive time intervals to obtain the state evolution loss value. This allows both instantaneous anomalies and chronic evolutionary deviations to be included in the training objective. Therefore, the gated recurrent network learns the equipment state evolution law rather than a single point error.

[0140] In detail, the initialization method of the device state history sequence is as follows: before the start of the first operating interval, the device state history sequence is initialized as a vector of all zero values, and this initial value is used as the input basis for the first linear mapping and the first gated cyclic network update.

[0141] In detail, the length setting method for multiple consecutive time intervals is as follows: training and updating are performed according to a sequence of running intervals of fixed length, and the most recent 8 to 64 consecutive running intervals are selected as a group of consecutive time intervals to take into account both short-term transitional changes and medium-term chronic evolution information.

[0142] In detail, the handling method for zero or excessively small predictor variance is as follows: First, recover the predictor variance using the natural logarithm of the predictor variance, and then ensure that the predictor variance is always greater than 0. When the recovered predictor variance is less than the smallest positive number that the system can represent, replace it with the smallest positive number before calculating the relative deviation value and the index probability loss value.

[0143] In detail, the online update sequence of the gated recurrent network is as follows: First, the predicted interval state vector, the mean of the predicted index, and the variance of the predicted index are calculated using the equipment state history sequence of the previous moment. Then, the interval state vector and the transition hysteresis area index of the current operating interval are read to calculate the state mean square error and the index probability loss value. Finally, the current interval state vector and the equipment state history sequence of the previous moment are input into the gated recurrent network to update and obtain the current equipment state history sequence.

[0144] Preferably, a comprehensive deviation is synthesized based on the deviation of the interval state vector and the deviation of the transition hysteresis area index, and an abnormal continuity assessment value is calculated using historical distribution probability, including: The mean square error between the integrated input tensor and the reconstructed output tensor is divided by the feature dimension of the integrated input tensor to obtain the dimension-normalized reconstruction error. The squared difference between the current transition hysteresis area index and the mean of the prediction index is divided by the sum of the variance of the prediction index and the numerical stability constant to obtain the corrected relative deviation value. The dimension-normalized reconstruction error, the state mean square error, and the corrected relative deviation value are added together to obtain the integrated deviation. The calculation formula is as follows: in, The comprehensive deviation amount, The dimension-uniformed reconstruction error, The mean square error is... Let be the feature dimension of the comprehensive input tensor. Let be the mean square error of the state. This is the current interval state vector. Let be the state vector of the prediction interval. The corrected relative deviation value, The current transition hysteresis area index, The mean of the predicted index, The variance of the prediction index, The numerical stability constant is denoted as . The total number of historical deviations greater than or equal to the comprehensive deviation in the historical interval is counted. This total number of historical deviations is then added to a single value and divided by the total number of sequences in the current operating interval to obtain the historical distribution probability. The calculation formula is as follows: in, The historical distribution probability is... The total number of sequences in the current running interval. The total number of historical deviations. For indicator functions, For the first The historical comprehensive deviation corresponding to each of the aforementioned historical intervals; Taking the natural logarithm of the historical distribution probability and then taking its negative value yields the anomaly continuous evaluation value, calculated using the following formula: in, The abnormal continuous evaluation value, It is the natural logarithm function.

[0145] The dimension-normalized reconstruction error is the normalized reconstruction error obtained by dividing the mean square error between the integrated input tensor and the reconstructed output tensor by the feature dimension of the integrated input tensor.

[0146] The feature dimension of the composite input tensor is the dimensionality of the composite input tensor used to represent different input features.

[0147] The corrected relative deviation is the normalized deviation obtained by dividing the square of the difference between the current transition hysteresis area index and the mean of the predicted index by the sum of the variance of the predicted index and the numerical stability constant.

[0148] The overall deviation is a single anomaly quantification result obtained by adding the dimension-normalized reconstruction error, the state mean square error, and the corrected relative deviation value.

[0149] The historical interval is the set of all running intervals that have been calculated before the current running interval.

[0150] The historical composite deviation is the composite deviation value corresponding to the historical interval.

[0151] The total historical deviation is the sum of the number of historical comprehensive deviations that are greater than or equal to the current comprehensive deviation within the historical interval.

[0152] The historical distribution probability is an empirical tail probability obtained based on the total number of historical deviations and the total number of sequences in the current operating interval.

[0153] The total number of sequences in the current running interval is the cumulative number of intervals from the first running interval to the current running interval.

[0154] The indicative function is used to determine whether the historical comprehensive deviation is greater than or equal to the current comprehensive deviation. It takes the value 1 when the condition is met and 0 when the condition is not met.

[0155] The continuous anomaly assessment value is a continuous anomaly intensity index obtained by taking the natural logarithm of the historical distribution probability and then taking the negative value.

[0156] In detail, because relying solely on the mean square error between the integrated input tensor and the reconstructed output tensor can only reflect whether the current operating interval deviates from the operating reference space, it cannot reflect the deviation of the interval state vector from historical evolution and the deviation of the transition hysteresis area index from the predicted distribution. Therefore, the comprehensive deviation is obtained by adding the dimension-normalized reconstruction error, the state mean square error, and the corrected relative deviation value. This can unify event space anomalies, state evolution anomalies, and index anomalies onto the same quantitative scale. Therefore, the comprehensive deviation is not an ordinary reconstruction error, but a synthetic representation of the three types of anomalies.

[0157] In detail, because the range of values ​​of the comprehensive deviation may vary greatly under different equipment, different load scenarios and different operating stages, a fixed threshold cannot be stably applied to all time periods. Therefore, by counting the total number of historical deviations greater than or equal to the comprehensive deviation in the historical interval, it is possible to determine the tail position of the current comprehensive deviation based on the historical distribution of the equipment itself.

[0158] In detail, if the total number of historical deviations is directly divided by the total number of sequences in the current operating range, the historical distribution probability will be too small and the initial fluctuations will be too large when the total number of historical deviations is 0. Therefore, adding the total number of historical deviations to a single value and then dividing by the total number of sequences in the current operating range can improve the numerical stability in the initial stage. Thus, the historical distribution probability is calculable under cold start conditions.

[0159] In detail, because the smaller the historical distribution probability, the more the current comprehensive deviation is at the tail of the history, and the negative value of the natural logarithm can map a very small probability to a larger positive number intensity, the continuous evaluation value of the anomaly can be obtained by taking the natural logarithm of the historical distribution probability and taking the negative value, which can transform the tail probability into a monotonically comparable continuous anomaly.

[0160] In detail, the statistical range of the historical interval is as follows: the baseline state set set at the beginning of system operation is used as the fixed historical interval. When the system adopts the online rolling update method, the most recent fixed number of continuous running intervals are forcibly limited to the historical interval. This is to keep the total number of sequences used to calculate the probability at present with a constant upper limit, avoid the logarithmic divergence of abnormal evaluation values ​​as the total running time increases, and the same statistical range rule must be maintained throughout the training and inference process.

[0161] In detail, the total number of sequences in the current running interval is determined during the cold start phase as follows: When processing the first running interval, the total number of sequences in the current running interval is set to 1, the historical distribution probability is set to 1, and the abnormal continuity evaluation value is set to 0. Starting from the second running interval, the historical distribution probability is calculated incrementally based on the accumulated number of running intervals.

[0162] In detail, the statistical method for the feature dimension of the composite input tensor is as follows: the total number of effective feature elements in the composite input tensor involved in the mean square error calculation is used as the feature dimension of the composite input tensor. When the composite input tensor consists of a normalized bus voltage sequence, a normalized load deviation sequence, and a transition hysteresis area index broadcast according to the time length, the feature dimension of the composite input tensor is equal to the effective time length multiplied by the number of features contained at each time position.

[0163] In detail, the online maintenance method for the total number of historical deviations is as follows: whenever a new comprehensive deviation is obtained, it is added to the historical comprehensive deviation sequence. When calculating the historical distribution probability in the next running interval, the relationship between all historical comprehensive deviations and the current comprehensive deviation is compared one by one. The comparison results that meet the condition of being greater than or equal to are accumulated to obtain the total number of historical deviations.

[0164] Preferably, the tap interval data is smoothed and corrected based on the abnormal continuous evaluation value to obtain a corrected load sequence, and the corrected load sequence is input into the basic load forecasting network to output the final demand forecast result, including: Between the starting boundary and the ending boundary of the interval, the historical distribution probability is used as a smoothing weight coefficient. The current active power sequence is multiplied by the smoothing weight coefficient, and the difference between this value and the smoothing weight coefficient is multiplied by the baseline demand mean. The products are then added together to obtain the corrected load sequence. The calculation formula is as follows: in, The smoothing weight coefficients are... The historical distribution probability is... For the corrected load sequence, For the current active power sequence, The benchmark mean of the demand is... For a specific moment in time; The historical revised load sequence and the external environment characteristic sequence for the future forecast period are input into the basic load forecasting network to output the final demand forecast result. The calculation formula is as follows: in, The final demand forecast result, The mapping function for the basic load forecasting network is... The corrected load sequence is the history from time one to the stated time. The external environment feature sequence from the stated time point to the future prediction period; Multiplying the absolute value of the abnormal continuous assessment value by the absolute value of the transition hysteresis area index yields the equipment abnormality contribution, calculated using the following formula: in, The amount contributed to the abnormality of the device. The abnormal continuous evaluation value, The absolute value of the transition hysteresis area index. The transition hysteresis area index is mentioned above.

[0165] The smoothing weighting factor is a weighting factor used to weight and merge the current active power series and the baseline demand mean.

[0166] The observed load weighting is the observed weighting result obtained by multiplying the current active power sequence by the smoothing weighting coefficient.

[0167] The baseline load weighted value is the result of multiplying the difference between a single value and a smoothing weighting coefficient by the baseline demand mean.

[0168] The corrected load sequence is the load sequence obtained by adding the observed load weights to the baseline load weights.

[0169] The historical corrected load sequence is the set of corrected load sequences that have undergone smoothing correction up to the current time point.

[0170] The future forecast period is the forecast step size of the final demand forecast result relative to the current time. A sampling step size of 1 to 96 is preferred. This range can cover ultra-short-term to intraday forecast scenarios. When the sampling period is 15 minutes, it corresponds to 15 minutes to 24 hours.

[0171] The mapping function of the basic load forecasting network is a forecasting function that maps the historical modified load sequence and the external environment characteristic sequence of the future forecast period to the final demand forecast result.

[0172] The final demand forecast result is the future demand forecast value output by the base load forecasting network after inputting the historical modified load sequence.

[0173] The absolute value of the transition hysteresis area index is the numerical value that retains only the magnitude of the transition hysteresis intensity after removing the direction sign.

[0174] The contribution of equipment anomalies is the intensity of the equipment anomaly impact obtained by multiplying the anomaly continuous assessment value by the absolute value of the transition hysteresis area index.

[0175] In detail, since a smaller historical distribution probability indicates that the current operating range deviates more from the normal historical distribution of the equipment, the historical distribution probability is used as a smoothing weight coefficient between the starting boundary and the ending boundary of the range. This allows operating ranges with high anomalies to rely more on the demand baseline mean, while operating ranges with low anomalies retain more of the current active power sequence.

[0176] In detail, because the current active power sequence may contain pseudo-demand components written by equipment, and the demand benchmark mean represents the normal demand center level that is not affected by voltage regulation, the current active power sequence and the demand benchmark mean are softly fused according to the smoothing weight coefficient to obtain a corrected load sequence, which can reduce pseudo-demand offsets in abnormal operating ranges while preserving the true load changes.

[0177] In detail, since the ultimate goal of anomaly detection is not only to issue alarms, but also to correct the input quality in the demand forecasting link, inputting the corrected load sequence into the basic load forecasting network and outputting the final demand forecasting result can directly feed the anomaly detection result back to the forecasting stage.

[0178] In detail, since the continuous anomaly assessment value reflects the anomaly intensity of the current operating range, and the absolute value of the transition hysteresis area index reflects the structural hysteresis intensity of the current operating range, multiplying the two to obtain the equipment anomaly contribution can jointly quantify the anomaly degree and structural characteristic intensity. Therefore, this quantity is more suitable for explaining the contribution of the current anomaly to the equipment side.

[0179] In detail, the smoothing weight coefficient is set as follows in the non-operational interval: when the current time is not between the start boundary and the end boundary of the interval, the smoothing weight coefficient is directly set to 1, so that the corrected load sequence is equal to the current active power sequence, thereby ensuring that the non-operational interval is not modified without reason.

[0180] In detail, the method for backfilling the historical corrected load sequence is as follows: After obtaining the corrected load sequence corresponding to each operating interval, the corrected load values ​​within that operating interval are written back to the full-time load data in their original chronological order. For time points that do not fall into any operating interval, the original value of the current active power sequence is directly retained. After completing the backfilling of all operating intervals, a continuous and complete historical corrected load sequence is obtained.

[0181] In detail, the future forecast period is set as follows: The future forecast period is determined based on the scheduling cycle and business span of the actual demand forecast task. When the task is a rolling ultra-short-term demand forecast, a sampling step size of 1 to 8 is preferred. When the task is an intraday demand forecast, a sampling step size of 24 to 96 is preferred. The definition of the future forecast period remains consistent throughout the entire training and deployment process.

[0182] In detail, the output method of equipment anomaly contribution is as follows: output the equipment anomaly contribution for each operating interval, and store each equipment anomaly contribution in a one-to-one correspondence with the anomaly continuous evaluation value, transition hysteresis area index and time range from the start boundary to the end boundary of the interval for subsequent use in equipment-side anomaly interpretation and operation and maintenance tracking.

[0183] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0184] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A deep learning-based time series anomaly detection system for energy demand forecasting, characterized in that, include: The operating interval division module is used to divide the operating interval by using the tap-out action of the on-load tap changer as the trigger point and the midpoint of the adjacent tap-out actions, and to obtain tap-out interval data containing the bus voltage sequence. The basic load forecasting module is used to construct the basic load forecasting network. During training, the data of the tap interval is masked to output a load deviation sequence that removes the influence of voltage regulation. The hysteresis index calculation module is used to calculate the transition hysteresis area index based on the bus voltage sequence and the load deviation sequence within the tap interval data. The operation mode learning module is used to input the tap interval data and the transition hysteresis area index into the conditional variational autoencoder, learn the operation reference space of the tap interval data and extract the interval state vector. The device status tracking module is used to input the interval status vector in chronological order to establish a historical sequence of device status. The anomaly quantification assessment module is used to synthesize a comprehensive deviation based on the deviation of the interval state vector and the deviation of the transition hysteresis area index, and to calculate the anomaly continuous assessment value through historical distribution probability. The demand forecasting correction module is used to smooth and correct the tap interval data based on the abnormal continuous evaluation value to obtain the corrected load sequence, and input the corrected load sequence into the basic load forecasting network to output the final demand forecasting result.

2. The deep learning time series anomaly detection system for energy demand forecasting according to claim 1, characterized in that, Using the tap changer's tapping action as the trigger point, the operating interval is divided using the midpoint of adjacent tapping actions to obtain tap interval data containing the bus voltage sequence, including: Obtain the tap position sequence, and determine the moment when the value changes in the tap position sequence as the trigger point; The median value of the time between the previous triggering time and the current triggering time is rounded down and used as the starting boundary of the running interval. The median value of the time between the current trigger time and the next trigger time is rounded down and used as the termination boundary of the running interval. The active power sequence, the bus voltage sequence, and the external environment feature sequence between the start boundary and the end boundary of the interval are extracted to jointly construct the tap interval data.

3. The deep learning time series anomaly detection system for energy demand forecasting according to claim 2, characterized in that, A basic load forecasting network is constructed, and the data of the tap changer interval is masked during training to output a load deviation sequence that removes the influence of voltage regulation, including: The historical active power sequence and the external environment feature sequence from history to the present are input into a temporal convolutional network to extract temporal context features; The temporal context features and the previous time-series cyclic hidden state vector are input into the gated cyclic unit to extract the current cyclic hidden state vector. The current cyclic hidden state vector is linearly mapped to obtain the demand baseline mean and demand baseline variance. Between the start boundary of the interval and the end boundary of the interval, the training masking coefficient is set to zero; in other time ranges, the training masking coefficient is set to one. The training shielding coefficient is used to calculate the loss of the current active power sequence, the mean of the demand baseline, and the variance of the demand baseline to obtain the basic network loss value and update the network parameters of the basic load prediction network. The load deviation sequence is obtained by subtracting the current active power sequence from the demand baseline mean.

4. The deep learning time series anomaly detection system for energy demand forecasting according to claim 3, characterized in that, Based on the bus voltage sequence and the load deviation sequence within the tap interval data, the transition hysteresis area index is calculated, including: Subtract the load deviation value corresponding to the starting boundary of the interval from the load deviation sequence to obtain the centered load deviation sequence; Subtracting the bus voltage value corresponding to the starting boundary of the interval from the bus voltage sequence yields a centered bus voltage sequence. The normalized load deviation sequence is obtained by dividing the centered load deviation sequence by the sum of the square root of the sum of squares between the starting boundary and the ending boundary of the interval and the numerical stability constant. The normalized bus voltage sequence is obtained by dividing the centralized bus voltage sequence by the sum of the square root of the sum of squares between the start boundary and the end boundary of the interval and the numerical stability constant. The normalized load deviation sequence and the normalized bus voltage sequence corresponding to the starting boundary of the interval are appended to the end of the sequence to form a closed endpoint; The arc length of the closed curve is calculated based on the adjacent differences between the normalized load deviation sequence and the adjacent differences between the normalized bus voltage sequence. The transition hysteresis area index is calculated using the discrete integral area formula based on the arc length of the closed curve, the adjacent mean values ​​of the normalized load deviation sequence, and the adjacent differences of the normalized bus voltage sequence.

5. The deep learning time series anomaly detection system for energy demand forecasting according to claim 4, characterized in that, The tap interval data and the transition hysteresis area index are input into a conditional variational autoencoder to learn the operating reference space of the tap interval data and extract the interval state vector, including: The normalized bus voltage sequence and the normalized load deviation sequence between the start boundary of the interval and the end boundary of the interval are spliced ​​together with the transition hysteresis area index broadcast according to the time length to obtain the comprehensive input tensor. The integrated input tensor is input into the encoding network of the conditional variational autoencoder to extract the state mean parameter and the state variance parameter; A standard normal distribution is preset and the random parameters of the standard normal distribution are obtained. Based on the state mean parameter, the state variance parameter and the random parameters of the standard normal distribution, reparameterized sampling is performed to extract the interval state vector. The interval state vector and the transition hysteresis area index are input into the decoding network of the conditional variational autoencoder to obtain the reconstructed output tensor. Based on the mean square error between the integrated input tensor and the reconstructed output tensor, and the relative entropy between the encoding distribution and the standard normal distribution, a spatial learning loss value is calculated to update the network parameters of the conditional variational autoencoder and learn the operating reference space.

6. The deep learning time series anomaly detection system for energy demand forecasting according to claim 5, characterized in that, The interval state vectors are input in chronological order to establish a device state history sequence, including: A linear mapping process is performed on the historical sequence of the device state at the previous moment to obtain the state vector of the prediction interval, the mean of the prediction index, and the variance of the prediction index. The current interval state vector and the previous device state history sequence are input into the gated recurrent network to update and obtain the current device state history sequence. Calculate the mean square error of the state between the current interval state vector and the predicted interval state vector; Calculate the squared difference between the current transition hysteresis area index and the mean of the predicted index, divide the squared difference by the variance of the predicted index to obtain the relative deviation value, and add the relative deviation value to the natural logarithm of the variance of the predicted index to obtain the index probability loss value. The state mean square error and the index probability loss value are accumulated over multiple consecutive time intervals to obtain the state evolution loss value, which is then used to update the network parameters of the gated recurrent network.

7. The deep learning time series anomaly detection system for energy demand forecasting according to claim 6, characterized in that, A comprehensive deviation is synthesized based on the deviation of the interval state vector and the deviation of the transition hysteresis area index, and an abnormal continuous evaluation value is calculated using historical distribution probability, including: The mean squared error between the integrated input tensor and the reconstructed output tensor is divided by the feature dimension of the integrated input tensor to obtain the dimension-normalized reconstruction error. The corrected relative deviation value is obtained by squared the difference between the current transition hysteresis area index and the mean of the predicted index, and divided by the sum of the variance of the predicted index and the numerical stability constant. The comprehensive deviation is obtained by adding the dimension-normalized reconstruction error, the state mean square error, and the corrected relative deviation value. The total number of historical deviations greater than or equal to the comprehensive deviation in the historical interval is counted. The total number of historical deviations is added to a value and then divided by the total number of sequences in the current operating interval to obtain the historical distribution probability. The anomaly continuous evaluation value is obtained by taking the natural logarithm of the historical distribution probability and then taking its negative value.

8. The deep learning time series anomaly detection system for energy demand forecasting according to claim 7, characterized in that, The tap interval data is smoothed and corrected based on the abnormal continuous assessment value to obtain a corrected load sequence. The corrected load sequence is then input into the basic load forecasting network to output the final demand forecast result, including: Between the starting boundary of the interval and the ending boundary of the interval, the historical distribution probability is used as a smoothing weighting coefficient; Multiply the current active power sequence by the smoothing weighting coefficient to obtain the observed load weighting value; The difference between a value and the smoothing weight coefficient is multiplied by the baseline demand mean to obtain the baseline load weighted value. The observed load weighted value is added to the baseline load weighted value to obtain the corrected load sequence; The historical revised load sequence and the external environment characteristic sequence for the future forecast period are input into the basic load forecasting network to output the final demand forecast result. The abnormal continuous evaluation value is multiplied by the absolute value of the transition hysteresis area index to obtain the equipment abnormality contribution.

Images