Distribution transformer weekly-ahead split-phase weight and overload early warning method considering three-phase load space-time characteristics and distribution transformer dynamic safety load domain

Abstract

The invention discloses a distribution transformer pre-cycle split-phase weight and overload early warning method considering three-phase load space-time characteristics and a distribution transformer dynamic safety load domain, which comprises the following steps of: 1) constructing a multi-channel characteristic graph based on the three-phase load space-time characteristics of a distribution transformer so as to represent an inter-phase load coupling relationship and a load multi-cycle characteristic; 2) establishing a distribution transformer pre-cycle split-phase load prediction model based on SwinLSTM-D; 3) constructing a double-weighted mixed loss function to optimize an attention distribution mode of the distribution transformer weekly front split-phase load prediction model; 4) depicting a distribution transformer dynamic safety load domain considering the distribution transformer hot spot temperature rise constraint so as to evaluate the distribution transformer limit load capacity; 5) solving the distribution transformer pre-cycle split-phase load prediction model to obtain a load prediction result; and comparing the load prediction result with the dynamic safe load domain of the distribution transformer, and generating distribution transformer weekly-ahead split-phase weight and overload early warning information. The method has the advantage of remarkable distribution transformer pre-cycle split-phase load prediction precision, and accurate distribution transformer pre-cycle split-phase weight and overload early warning can be carried out.

Description

Technical Field

[0001] The present invention relates to the field of warning of the overloaded operation state of distribution transformers, and specifically to a method for predicting the phase-by-phase overload of distribution transformers in a weekly cycle considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers. Background Art

[0002] With the large-scale access of new power sources and loads, the spatio-temporal characteristics of the three-phase loads of distribution transformers have changed profoundly, and the high-incidence of three-phase unbalanced working conditions has ultimately led to the mixed occurrence of single-phase, two-phase, and three-phase overload problems of distribution transformers, endangering the safety of the power grid. Accurately predicting the phase-by-phase overload events of distribution transformers in a weekly cycle can reserve sufficient time for the formulation and implementation of governance strategies, which is of great significance for eliminating potential accidents and enhancing the safety and stability of the power grid.

[0003] The existing methods for predicting the overload of transformers mainly include traditional warning methods represented by classification models and new warning methods represented by deep learning. Among them, the traditional methods only issue warning results based on external factors such as temperature, week, and power supply category, and have defects such as poor interpretability and low accuracy. The new warning methods comprehensively consider the influence of various internal and external factors and have high warning accuracy, mainly consisting of two parts: load forecasting and evaluation of the load capacity of transformers.

[0004] However, the current load forecasting methods mostly target the overall load forecasting, and it is difficult to fully explore the dynamic operation laws of the three-phase loads of distribution transformers, and it is impossible to ensure the accuracy of the phase-by-phase load forecasting in a weekly cycle. In addition, since distribution transformers are more susceptible to the influence of the three-phase unbalance degree under light load (load rate ≤ 30%) and heavy overload (load rate ≥ 80%) working conditions, which in turn leads to an abnormal increase in losses. This requires that the phase-by-phase load forecasting model of distribution transformers in a weekly cycle has good performance under these two working conditions. However, the current methods have not met this requirement.

[0005] At the same time, the dynamic transformer rating (DTR) is defined as the maximum load that a transformer can withstand stably under the change of load and ambient temperature. The existing evaluation models do not consider the influence of the three-phase unbalanced working conditions on the load capacity of distribution transformers, and the current overload warning methods only issue warning results based on the proportion of the load in the rated capacity, ignoring the difference between the load capacity of the distribution transformer and the rated capacity.

[0006] The above technical defects have led to the basic failure of the existing new methods for predicting the overload of distribution transformers. Therefore, there is an urgent need for a method for predicting the phase-by-phase overload of distribution transformers in a weekly cycle considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers. Summary of the Invention

[0007] The purpose of the present invention is to provide a method for predicting the phase-by-phase overload of distribution transformers in a weekly cycle considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers, including the following steps:

[0008] 1) Based on the spatio-temporal characteristics of the three-phase load of the distribution transformer, construct a multi-channel feature map to characterize the inter-phase load coupling relationship and the multi-periodic characteristics of the load;

[0009] 2) Establish a pre-week phase-separated load prediction model for the distribution transformer based on SwinLSTM-D;

[0010] 3) Construct a double-weighted hybrid loss function to optimize the attention allocation mode of the pre-week phase-separated load prediction model for the distribution transformer;

[0011] 4) Characterize the dynamic safety load domain of the distribution transformer considering the temperature rise constraint of the hot spot of the distribution transformer to evaluate the ultimate load capacity of the distribution transformer;

[0012] 5) Solve the pre-week phase-separated load prediction model for the distribution transformer to obtain the load prediction result; compare the load prediction result with the dynamic safety load domain of the distribution transformer to generate the pre-week phase-separated overload warning information for the distribution transformer.

[0013] Furthermore, in step 1), the steps of constructing a multi-channel feature map to characterize the inter-phase load coupling relationship and the multi-periodic characteristics of the load include:

[0014] 1.1) Based on the phase-X load data of the distribution transformer X = (x 1 , x 2 , …, x n ) and the phase-Y load data Y = (y 1 , y 2 , …, y n ), construct the data set D = {(x i , y i ), i = 1, 2, …, n}; n is the sequence length;

[0015] Among them, the mutual information number of the phase-X load data and the phase-Y load data is as follows:

[0016]

[0017] In the formula: I MI (X; Y) is the mutual information coefficient between phase-X and phase-Y; p(x, y) is the joint probability density function of phase-X and phase-Y; p(x) and p(y) are the marginal probability density functions of phase-X and phase-Y respectively;

[0018] 1.2) Discretize the data set D on a two-dimensional plane and divide it into a × b grids, denoted as grid G, and calculate the maximum mutual information I*MI(D∣ G , a, b);

[0019] The maximum mutual information I*MI(D∣ G , a, b) is as follows:

[0020]

[0021] Where: D∣ G is the probability distribution that the points in D fall into G; I MI (D∣ G , a, b) is the estimated value of the mutual information between the X phase and the Y phase under the probability distribution D∣ G ;

[0022] 1.3) Calculate the maximum information coefficient I MIC (X; Y), that is:

[0023]

[0024] Where: B(n) is the constraint function that determines the maximum number of grids;

[0025] 1.4) Based on the maximum information coefficient I MIC (X; Y), quantify the coupling degree between loads in different seasons respectively;

[0026] 1.5) For each phase load, select the historical load data of this phase and the historical load data of other phases with I MIC (X; Y) greater than the preset threshold as input features;

[0027] 1.6) For each phase in the input features, encode the single-phase time-series load sequence into a single-channel two-dimensional feature map, so that the processing method of the prediction network for load information is upgraded from line to surface;

[0028] 1.7) Independently construct the load data of each phase in the input features into adjacent channels, and feature interaction and fusion occur between the channels to form a spatial association pattern;

[0029] 1.8) Repeat step 1.7), encode the load data of the input features into a multi-channel feature map, and independently construct the load of each phase into adjacent channels; among them, the coupling relationship between the phase loads is reflected as the spatial dependence relationship of the corresponding regions of the adjacent channels, and each channel contains the multi-periodic characteristics of the phase loads expressed by spatial features.

[0030] Furthermore, the maximum information coefficient I MIC (X; Y) is used to quantify the coupling degree between three-phase and overall loads; I MIC (X; Y) ranges from [0, 1]. When the I MIC (X; Y) value between two phases is larger, the coupling degree is deeper.

[0031] Furthermore, the input data of the distribution transformer weekly pre-phase load prediction model based on SwinLSTM-D is a multi-channel feature map, and the output data is the predicted value of the phase load.

[0032] Further, in step 3), the double weighted mixed loss function DWH is as follows:

[0033]

[0034] where: N represents the total number of samples, A i is the actual load value, F i is the predicted value corresponding to A i , ω is the weight, which can determine the attention degree of the model to the light load condition, τ 1 and τ 2 are the function segmentation thresholds, which are used to define the load ranges of the light load and heavy overload conditions. Both ω and τ 1 and τ 2 are hyperparameters.

[0035] Further, in step 4), the steps of characterizing the distribution transformer dynamic safety load domain considering the distribution transformer hot spot temperature rise constraint include:

[0036] 4.1) Adopt a distribution transformer hot spot temperature calculation model considering the three-phase unbalanced condition to solve the hot spot temperatures of the windings and core of phases A, B, and C of the distribution transformer, as well as the top oil temperature in the next week;

[0037] 4.2) Based on the weekly prediction data of the hot spot temperature and the top oil temperature, characterize the distribution transformer dynamic safety load domain considering the distribution transformer hot spot temperature rise constraint to evaluate the distribution transformer's ultimate load capacity.

[0038] Further, the distribution transformer hot spot temperature calculation model is as follows:

[0039]

[0040] where: θ amb is the ambient temperature, R oil is the heat transfer thermal resistance between the transformer oil and the external environment, R a , R b , R c are the heat transfer thermal resistances between the three-phase windings and the transformer oil respectively, C oil , C a , C b , C c , C fe are the heat capacities of the transformer oil, three-phase windings, and core respectively, q fe is the core loss, θ hs,a , θ hs,b , θ hs,c , θ fe , θ to are the hot spot temperatures of the windings of phases A, B, and C and the core, as well as the top oil temperature, q cu,a , q cu,b , q cu,care the load losses of the three-phase windings respectively;

[0041] Among them, the load loss q cu,i is as follows:

[0042] q cu,i = (K i (I N )) i ) 2 R i , i ∈ {a, b, c} (7)

[0043] In the formula: K i is the phase load rate, (I N ) i is the phase rated current, and R i is the sum of the resistance value of the high-voltage winding reduced to the low-voltage side and the resistance of the low-voltage winding.

[0044] Furthermore, in step 4.2), the steps of characterizing the dynamic safety load domain of the distribution transformer considering the hot-spot temperature rise constraint include:

[0045] 4.2.1) Collect the thermal and electrical parameters of the target distribution transformer and obtain the weekly-ahead phase-by-phase load prediction results;

[0046] 4.2.2) Establish a calculation model for the hot-spot temperature of the distribution transformer based on the thermoelectric analogy theory;

[0047] 4.2.3) Based on the weekly-ahead phase-by-phase load prediction results of the distribution transformer, continuously solve the thermal circuit model by dynamically adjusting the starting temperature, and calculate the weekly-ahead temperature change curves of the hot-spot temperatures of the three-phase windings and the core and the top oil temperature;

[0048] 4.2.4) Follow the distribution transformer design specifications and safety standards to stipulate the hot-spot temperature rise constraint and the safe operation time limit.

[0049] 4.2.5) For each moment of the weekly-ahead temperature prediction curve, simulate the temperature change process of the distribution transformer within the safe operation time limit under different load scenarios one by one, identify and define the limit load space of the distribution transformer that meets the hot-spot temperature rise constraint, and complete the characterization of the dynamic safety load domain of the distribution transformer.

[0050] Furthermore, in step 5), before generating the weekly-ahead phase-by-phase overload warning information of the distribution transformer, set the boundary of the dynamic safety load domain of the distribution transformer as the overload warning boundary, and 80% of the overload warning boundary is the heavy-load warning boundary.

[0051] Furthermore, if the load prediction result reaches the overload warning boundary, the generated weekly-ahead phase-by-phase overload warning information of the distribution transformer is an overload condition warning;

[0052] If the load prediction result reaches the heavy-load warning boundary, the generated weekly-ahead phase-by-phase overload warning information of the distribution transformer is a heavy-load condition warning.

[0053] The technical effect of the present invention is beyond doubt. The present invention has a significant advantage in the prediction accuracy of the pre-weekly phase-separated load of distribution transformers, and can perform accurate pre-weekly phase-separated overload and heavy-load early warnings for distribution transformers. Description of the Drawings

[0054] Figure 1 is a flow chart of a pre-weekly phase-separated overload and heavy-load early warning method for distribution transformers considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers;

[0055] Figure 2 is a flow chart for characterizing the dynamic safety load domain of distribution transformers considering the temperature rise constraint of the hot spot of the distribution transformer;

[0056] Figure 3 is a pre-weekly phase-separated overload and heavy-load early warning strategy diagram for distribution transformers based on the dynamic safety load domain of distribution transformers. Detailed Embodiments

[0057] The present invention will be further described below in conjunction with embodiments, but it should not be understood that the above-mentioned subject scope of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes made according to ordinary technical knowledge and customary means in the art should be included within the protection scope of the present invention.

[0058] Embodiment 1:

[0059] Refer to Figures 1 to 3 , a pre-weekly phase-separated overload and heavy-load early warning method for distribution transformers considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers, includes the following steps:

[0060] 1) Based on the spatio-temporal characteristics of the three-phase load of the distribution transformer, construct a multi-channel feature map to characterize the inter-phase load coupling relationship and the multi-periodic characteristics of the load, so as to transform the coupling relationship between the phase loads in the feature combination and the multi-periodic characteristics of the load sequence into a spatial feature representation, enhance the effectiveness of the spatio-temporal characteristics expression of the three-phase load of the distribution transformer, and improve the prediction accuracy;

[0061] 2) Establish a pre-weekly phase-separated load prediction model for distribution transformers based on SwinLSTM-D;

[0062] 3) Construct a dual-weighted hybrid loss function to optimize the attention allocation mode of the pre-weekly phase-separated load prediction model for distribution transformers;

[0063] 4) Characterize the dynamic safety load domain of the distribution transformer considering the temperature rise constraint of the hot spot of the distribution transformer to evaluate the ultimate load capacity of the distribution transformer;

[0064] 5) Solve the pre-weekly phase-separated load prediction model for distribution transformers to obtain the load prediction result; compare the load prediction result with the dynamic safety load domain of the distribution transformer to generate pre-weekly phase-separated overload and heavy-load early warning information.

[0065] In step 1), the steps of constructing a multi-channel feature map to characterize the inter-phase load coupling relationship and the load multi-periodicity characteristics include:

[0066] 1.1) Based on the X-phase load data of the distribution transformer X = (x 1 , x 2 , …, x n ) and the Y-phase load data Y = (y 1 , y 2 , …, y n ), construct the data set D = {(x i , y i ), i = 1, 2, …, n}; n is the sequence length;

[0067] Among them, the mutual information numbers of the X-phase load data and the Y-phase load data are as follows:

[0068]

[0069] In the formula: I MI (X; Y) is the mutual information coefficient between the X-phase and the Y-phase; p(x, y) is the joint probability density function of the X-phase and the Y-phase; p(x) and p(y) are the marginal probability density functions of the X-phase and the Y-phase respectively;

[0070] 1.2) Discretize the data set D on a two-dimensional plane and divide it into a × b grids, denoted as grid G, and calculate the maximum mutual information I*MI(D∣ G , a, b) under different grid partitions;

[0071] The maximum mutual information I*MI(D∣ G , a, b) is as follows:

[0072]

[0073] In the formula: D∣ G is the probability distribution of the points in D falling into G; I MI (D∣ G , a, b) is the estimated mutual information value between the X-phase and the Y-phase under the probability distribution D∣ G ;

[0074] 1.3) Calculate the maximum information coefficient I MIC (X; Y), that is:

[0075]

[0076] In the formula: B(n) is the constraint function that determines the maximum number of grids;

[0077] 1.4) Based on the maximum information coefficient I MIC(X; Y) quantifies the coupling degree between loads in different seasons respectively;

[0078] 1.5) For each phase load, select the historical load data of this phase and the historical load data of other phases with I MIC (X; Y) greater than the preset threshold as input features;

[0079] 1.6) For each phase in the input features, encode the single-phase time-series load sequence into a single-channel two-dimensional feature map, so that the processing method of the prediction network for load information is upgraded from line to surface;

[0080] 1.7) Independently construct the load data of each phase in the input features into adjacent channels, and feature interaction and fusion occur between the channels to form a spatial association pattern;

[0081] 1.8) Repeat step 1.7), encode the load data of the input features into a multi-channel feature map, and independently construct the load of each phase into adjacent channels; among them, the coupling relationship between phases is reflected as the spatial dependence relationship of the corresponding regions of adjacent channels, and each channel contains the multi-periodic characteristics of the load of each phase expressed by spatial features.

[0082] The maximum information coefficient I MIC (X; Y) is used to quantify the coupling degree between three-phase and overall loads; I MIC (X; Y) ranges from [0, 1]. When the I MIC (X; Y) value between two phases is larger, the coupling degree is deeper.

[0083] The input data of the distribution transformer weekly pre-phase load prediction model based on SwinLSTM-D is a multi-channel feature map, and the output data is the predicted value of the phase load.

[0084] In step 3), the double weighted hybrid loss function DWH is as follows:

[0085]

[0086] In the formula: N represents the total number of samples, A i is the actual load value, F i is the predicted value corresponding to A i , ω is the weight, which can determine the attention degree of the model to the light load condition, τ 1 and τ 2 are the function segmentation thresholds, used to define the load range of the light load and heavy overload conditions, and ω and τ 1 and τ 2 are all hyperparameters.

[0087] In step 4), the steps to depict the distribution transformer dynamic safety load domain considering the distribution transformer hot spot temperature rise constraint include:

[0088] 4.1) Adopt a calculation model for the hot-spot temperature of distribution transformers considering three-phase unbalanced operating conditions to solve the hot-spot temperatures of the A, B, and C phase windings, the core, and the top oil temperature of the distribution transformer in the next week.

[0089] 4.2) Based on the weekly prediction data of the hot-spot temperature and the top oil temperature, characterize the dynamic safety load domain of the distribution transformer considering the hot-spot temperature rise constraint of the distribution transformer to evaluate the ultimate load capacity of the distribution transformer.

[0090] The calculation model for the hot-spot temperature of the distribution transformer is as follows:

[0091]

[0092] In the formula: θ amb is the ambient temperature, R oil is the heat transfer resistance between the transformer oil and the external environment, R a , R b , R c are the heat transfer resistances between the three-phase windings and the transformer oil respectively, C oil , C a , C b , C c , C fe are the heat capacities of the transformer oil, the three-phase windings, and the core respectively, q fe is the core loss, θ hs,a , θ hs,b , θ hs,c , θ fe , θ to are the hot-spot temperatures of the A, B, and C phase windings and the core, and the top oil temperature respectively, q cu,a , q cu,b , q cu,c are the load losses of the three-phase windings respectively;

[0093] Among them, the load loss q cu,i is as follows:

[0094] q cu,i =(K i (I N ) i ) 2 R i , i∈{a, b, c} (7)

[0095] In the formula: K i is the phase load rate, (I N ) i is the phase rated current, R i is the sum of the high-voltage winding resistance reduced to the low-voltage side and the low-voltage winding resistance.

[0096] In step 4.2), the steps to characterize the dynamic safety load domain of the distribution transformer considering the hot-spot temperature rise constraint of the distribution transformer include:

[0097] 4.2.1) Collect the thermal and electrical parameters of the target distribution transformer and obtain the weekly pre-phase load forecasting results;

[0098] 4.2.2) Establish a calculation model for the hot spot temperature of the distribution transformer based on the thermoelectric analogy theory;

[0099] 4.2.3) Based on the weekly pre-phase load forecasting results of the distribution transformer, continuously solve the thermal circuit model by dynamically adjusting the starting temperature, and calculate the weekly pre-temperature change curves of the hot spot temperatures of the three-phase windings and the iron core and the top oil temperature;

[0100] 4.2.4) Follow the distribution transformer design specifications and safety standards to stipulate the hot spot temperature rise constraints and the safe operation time limit.

[0101] 4.2.5) For each moment of the weekly pre-temperature forecasting curve, simulate the temperature change process of the distribution transformer within the safe operation time limit under different load scenarios one by one, identify and define the limit load space of the distribution transformer that meets the hot spot temperature rise constraints, and complete the characterization of the dynamic safety load domain of the distribution transformer.

[0102] In step 5), before generating the weekly pre-phase overload and heavy overload warning information of the distribution transformer, set the boundary of the dynamic safety load domain of the distribution transformer as the overload warning boundary, and 80% of the overload warning boundary is the heavy overload warning boundary.

[0103] If the load forecasting result reaches the overload warning boundary, the generated weekly pre-phase overload and heavy overload warning information of the distribution transformer is the overload condition warning;

[0104] If the load forecasting result reaches the heavy overload warning boundary, the generated weekly pre-phase overload and heavy overload warning information of the distribution transformer is the heavy overload condition warning.

[0105] Example 2:

[0106] A weekly pre-phase overload and heavy overload warning method for a distribution transformer considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of the distribution transformer, including the following steps:

[0107] 1) Based on the spatio-temporal characteristics of the three-phase loads of the distribution transformer, construct a multi-channel feature map to characterize the inter-phase load coupling relationship and the load multi-periodicity characteristics;

[0108] 2) Establish a weekly pre-phase load forecasting model for the distribution transformer based on SwinLSTM-D;

[0109] 3) Construct a double weighted hybrid loss function to optimize the attention allocation mode of the weekly pre-phase load forecasting model of the distribution transformer;

[0110] 4) Characterize the dynamic safety load domain of the distribution transformer considering the hot spot temperature rise constraints of the distribution transformer to evaluate the limit load capacity of the distribution transformer;

[0111] 5) Solve the phase - separated load prediction model for the distribution transformer a week in advance to obtain the load prediction results; compare the load prediction results with the dynamic safety load domain of the distribution transformer to generate the phase - separated overload warning information for the distribution transformer a week in advance.

[0112] Embodiment 3:

[0113] For the phase - separated overload warning method of the distribution transformer considering the spatio - temporal characteristics of three - phase loads and the dynamic safety load domain of the distribution transformer, the technical content is the same as that of Embodiment 2. Further, in step 1), the steps of constructing a multi - channel feature map to characterize the inter - phase load coupling relationship and the load multi - periodicity characteristics include:

[0114] 1.1) Based on the phase - X load data of the distribution transformer \(X=(x 1 ,x 2 ,…,x n )\) and the phase - Y load data \(Y=(y 1 ,y 2 ,…,y n )\), construct the data set \(D =\{(x i ,y i ),i = 1,2,…,n\}\); \(n\) is the sequence length;

[0115] Among them, the mutual information number of the phase - X load data and the phase - Y load data is as follows:

[0116]

[0117] In the formula: \(I MI (X;Y)\) is the mutual information coefficient between phase - X and phase - Y; \(p(x,y)\) is the joint probability density function of phase - X and phase - Y; \(p(x)\) and \(p(y)\) are the marginal probability density functions of phase - X and phase - Y respectively;

[0118] 1.2) Discretize the data set \(D\) on a two - dimensional plane and divide it into \(a×b\) grids, denoted as grid \(G\), and calculate the maximum mutual information \(I*MI(D∣ G ,a,b)\) under different grid partitions;

[0119] The maximum mutual information \(I*MI(D∣ G ,a,b)\) is as follows:

[0120]

[0121] In the formula: \(D∣ G \) is the probability distribution of the points in \(D\) falling into \(G\); \(I MI (D∣ G ,a,b)\) is the estimated value of the mutual information between phase - X and phase - Y under the probability distribution \(D∣ G ;

[0122] 1.3) Calculate the maximum information coefficient \(I MIC(X; Y), i.e.,

[0123]

[0124] where: B(n) is a constraint function that determines the maximum number of grids;

[0125] 1.4) Based on the maximum information coefficient I MIC (X; Y), quantify the coupling degree between loads in different seasons respectively;

[0126] 1.5) For each phase of load, select the historical load data of this phase and the historical load data of other phases where I MIC (X; Y) is greater than the preset threshold as input features;

[0127] 1.6) For each phase in the input features, encode the single-phase time-series load sequence into a single-channel two-dimensional feature map, so that the processing method of the prediction network for load information is upgraded from line to surface;

[0128] 1.7) Independently construct the load data of each phase in the input features into adjacent channels, and feature interaction and fusion occur between the channels to form a spatial correlation pattern;

[0129] 1.8) Repeat step 1.7), encode the load data of the input features into a multi-channel feature map, and independently construct the load of each phase into adjacent channels; among them, the coupling relationship between phases is reflected as the spatial dependence relationship of the corresponding regions of adjacent channels, and each channel contains the multi-periodic characteristics of the load of each phase expressed by spatial features.

[0130] Example 4:

[0131] A method for predicting pre-week phase-by-phase heavy overload of distribution transformers considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers. The technical content is the same as any one of Examples 2-3. Further, the maximum information coefficient I MIC (X; Y) is used to quantify the coupling degree between three-phase and overall loads; I MIC (X; Y) ranges from [0, 1]. When the I MIC (X; Y) value between two phases is larger, the coupling degree is deeper.

[0132] Example 5:

[0133] A method for predicting pre-week phase-by-phase heavy overload of distribution transformers considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers. The technical content is the same as any one of Examples 2-4. Further, the input data of the pre-week phase-by-phase load prediction model based on SwinLSTM-D is a multi-channel feature map, and the output data is the predicted value of the phase-by-phase load.

[0134] Example 6:

[0135] A method for predicting weekly pre - phase - separated heavy and overload of distribution transformers considering the spatio - temporal characteristics of three - phase loads and the dynamic safety load domain of distribution transformers. The technical content is the same as any one of Embodiments 2 - 5. Further, in step 3), the double - weighted hybrid loss function is as follows:

[0136]

[0137] In the formula: N represents the total number of samples, A i is the actual load value, F i is the predicted value corresponding to A i , ω is the weight, which can determine the degree of attention of the model to the light - load condition, τ 1 and τ 2 are the function - segment thresholds used to define the load ranges of the light - load and heavy - overload conditions. Both ω and τ 1 and τ 2 are hyperparameters.

[0138] Embodiment 7:

[0139] A method for predicting weekly pre - phase - separated heavy and overload of distribution transformers considering the spatio - temporal characteristics of three - phase loads and the dynamic safety load domain of distribution transformers. The technical content is the same as any one of Embodiments 2 - 6. Further, in step 4), the steps of characterizing the dynamic safety load domain of the distribution transformer considering the hot - spot temperature rise constraint of the distribution transformer include:

[0140] 4.1) Adopt a calculation model for the hot - spot temperature of the distribution transformer considering the three - phase unbalanced condition to solve the hot - spot temperatures of the windings and the core of phases A, B, and C of the distribution transformer and the top - layer oil temperature in the next week;

[0141] 4.2) Based on the weekly pre - prediction data of the hot - spot temperature and the top - layer oil temperature, characterize the dynamic safety load domain of the distribution transformer considering the hot - spot temperature rise constraint to evaluate the ultimate load - carrying capacity of the distribution transformer.

[0142] Embodiment 8:

[0143] A method for predicting weekly pre - phase - separated heavy and overload of distribution transformers considering the spatio - temporal characteristics of three - phase loads and the dynamic safety load domain of distribution transformers. The technical content is the same as any one of Embodiments 2 - 7. Further, the calculation model for the hot - spot temperature of the distribution transformer is as follows:

[0144]

[0145] In the formula: θ amb is the ambient temperature, R oil is the heat - transfer thermal resistance between the transformer oil and the external environment, R a , R b , R c are the heat - transfer thermal resistances between the three - phase windings and the transformer oil respectively, C oil , C a , C b , Cc , C fe are the heat capacities of transformer oil, three-phase windings, and iron core respectively, q fe is the iron core loss, θ hs,a , θ hs,b , θ hs,c , θ fe , θ to are the hot spot temperatures of the three-phase windings and the iron core, and the top oil temperature respectively, q cu,a , q cu,b , q cu,c are the load losses of the three-phase windings respectively;

[0146] Among them, the load loss q cu,i is as follows:

[0147] q cu,i =(K i (I N ) i ) 2 R i , i ∈ {a, b, c} (7)

[0148] In the formula: K i is the phase load rate, (I N ) i is the phase rated current, R i is the sum of the high-voltage winding resistance reduced to the low-voltage side and the low-voltage winding resistance.

[0149] Example 9:

[0150] A method for predicting pre-week phase-separated heavy overload of distribution transformers considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers. The technical content is the same as any one of Examples 2-8. Further, in step 4.2), the steps of characterizing the dynamic safety load domain of distribution transformers considering the hot spot temperature rise constraint of distribution transformers include:

[0151] 4.2.1) Collect the thermal and electrical parameters of the target distribution transformer and obtain the pre-week phase-separated load prediction results;

[0152] 4.2.2) Establish a calculation model for the hot spot temperature of the distribution transformer based on the thermoelectric analogy theory;

[0153] 4.2.3) Based on the pre-week phase-separated load prediction results of the distribution transformer, continuously solve the thermal circuit model by dynamically adjusting the starting temperature, and calculate the pre-week temperature change curves of the hot spot temperatures and the top oil temperature of the three-phase windings and the iron core;

[0154] 4.2.4) Follow the design specifications and safety standards of the distribution transformer, and stipulate the hot spot temperature rise constraint and the safe operation time limit.

[0155] 4.2.5) For each moment of the weekly temperature prediction curve, simulate the temperature change process of the distribution transformer within the safe operation time limit under different load scenarios one by one, identify and define the limit load space of the distribution transformer that meets the hot spot temperature rise constraint, and complete the characterization of the dynamic safety load domain of the distribution transformer.

[0156] Example 10:

[0157] A method for predicting weekly phase-separated overload of a distribution transformer considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of the distribution transformer, the technical content is the same as any one of Examples 2-9. Further, in step 5), before generating the weekly phase-separated overload warning information of the distribution transformer, set the boundary of the dynamic safety load domain of the distribution transformer as the overload warning boundary, and 80% of the overload warning boundary is the heavy load warning boundary.

[0158] Example 11:

[0159] A method for predicting weekly phase-separated overload of a distribution transformer considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of the distribution transformer, the technical content is the same as any one of Examples 2-10. Further, if the load prediction result reaches the overload warning boundary, the generated weekly phase-separated overload warning information of the distribution transformer is an overload condition warning;

[0160] If the load prediction result reaches the heavy load warning boundary, the generated weekly phase-separated overload warning information of the distribution transformer is a heavy load condition warning.

[0161] Example 12:

[0162] A method for predicting weekly phase-separated overload of a distribution transformer considering the spatio-temporal characteristics of three-phase loads and the dynamic safety load domain of the distribution transformer mainly includes the following steps:

[0163] 1. Explore the spatio-temporal characteristics of the three-phase load of the distribution transformer, and establish a feature enhanced expression method based on the dynamic selection technology of feature combination and the spatial representation technology. The main steps are as follows:

[0164] 1) Deeply explore the dynamic operation law of the three-phase load of the distribution transformer, and summarize the spatio-temporal characteristics as follows:

[0165] First, the mixed use of single-phase and three-phase power supply methods in the distribution transformer substation area results in a coupling relationship between the three-phase and overall loads of the distribution transformer, and the seasonal change of the coupling degree caused by the seasonal alternation of the fluctuation laws of various new energy sources and loads leads to a seasonal differential coupling relationship between the three-phase and overall loads of the distribution transformer.

[0166] In addition, the load sequence shows complex multi-periodic characteristics, covering short periods (such as intra-day fluctuations), medium periods (such as daily repeated patterns), and long periods (such as weekly cycles, etc.). The various period dimensions are intertwined with each other and jointly affect the current load level.

[0167] Finally, as the overall load of the distribution transformer is the superposition of three-phase loads, due to the phase and amplitude differences in the fluctuations of the three-phase loads, the superposition effect will obscure the differential fluctuation characteristics of the three-phase loads, resulting in the cancellation or enhancement of some fluctuations in the overall load.

[0168] In summary, fully capturing the seasonal differential coupling relationship of the three-phase and overall loads of the distribution transformer and the multi-periodic characteristics of the load sequence are important prerequisites for improving the prediction accuracy of the weekly pre-split phase load of the distribution transformer. At the same time, the superposition effect between the three-phase load and the overall load obscures the differential fluctuation characteristics of the three-phase load, making it impossible to accurately infer the three-phase load based on the overall load prediction.

[0169] 2) Construct a dynamic selection technology for feature combinations considering the seasonal differential coupling relationship of three-phase loads, enabling the model to focus on key features to improve training efficiency, as follows:

[0170] It should be noted that the overall load is usually a concentrated manifestation of the change trend of the three-phase load. Therefore, it is included in the range of candidates for the dynamic selection of feature combinations, and it is determined whether to be used as an input feature based on the quantification result of the coupling degree, but it has been clearly defined as not being the prediction target.

[0171] For the dataset D = {(x 1 ,x 2 ,…,x n ) and Y-phase load data Y = (y 1 ,y 2 ,…,y n )} composed of the X-phase load data of the distribution transformer X = (x i ,y i ), where i = 1, 2, …, n and n is the sequence length, the calculation model of its mutual information number is expressed as:

[0172]

[0173] In the formula: I MI (X; Y) is the mutual information coefficient between the X phase and the Y phase; p(x, y) is the joint probability density function of the X phase and the Y phase; p(x) and p(y) are the marginal probability density functions of the X phase and the Y phase, respectively.

[0174] First, discretize D on a two-dimensional plane and divide it into a × b grids, denoted as G. Denote the probability distribution of the points in D falling into G as D∣ G , and denote I MI (D∣ G ,a,b) as the estimated value of the mutual information between the X phase and the Y phase in this partitioning case. Then, select different a and b and calculate multiple times to obtain the maximum mutual information I*MI(D∣ G ,a,b) under different grid partitions, which is expressed as:

[0175]

[0176] Finally, the maximum information coefficient is obtained and expressed as:

[0177]

[0178] In the formula: B(n) is a constraint function that determines the maximum number of grids. Generally, B(n) = n 0.6 .

[0179] Based on equations (1)-(3), the coupling degrees between the three phases and the overall load can be quantified respectively. The value range of I MIC (X; Y) is [0, 1]. When the value of I MIC (X; Y) between two phases is larger, the coupling degree is deeper. On the contrary, when the value of I MIC (X; Y) is smaller, the coupling degree between the two phases is shallower.

[0180] In view of the seasonal differences in the coupling relationship between the three phases and the overall load, the coupling degree between the loads is quantified for each season separately. For each phase of the load, only its historical data and the historical data of other phases with I MIC (X; Y) values exceeding the threshold are selected as input features to eliminate redundancy and optimize the model training efficiency. The threshold is a hyperparameter.

[0181] 3) A spatial representation technology for the coupling relationship between inter-phase loads and the multi-periodic characteristics of loads is proposed to enhance the effectiveness of expressing the spatio-temporal characteristics of the three-phase load of the distribution transformer and improve the prediction accuracy, as follows:

[0182] First, for each phase in the feature combination, the single-phase time-series load sequence is encoded into a single-channel two-dimensional feature map respectively, so that the processing method of the prediction network for the load information is upgraded from one-dimensional to two-dimensional. At this time, the prediction network does not need to crawl the load sequence multiple times, and can completely capture the multi-periodic characteristics of the load sequence expressed by spatial features within a single time step.

[0183] Second, the load data of each phase in the feature combination are independently constructed into adjacent channels. As the network deepens, feature interaction and fusion will occur between the channels, forming rich spatial association patterns. Based on this, the model can efficiently capture the coupling relationship between the loads of each phase.

[0184] Finally, the load data of the feature combination are encoded into a multi-channel feature map, and the load of each phase is independently constructed into adjacent channels. At this time, the coupling relationship between the inter-phase loads is reflected as the spatial dependence relationship of the corresponding regions of the adjacent channels. Each channel contains the multi-periodic characteristics of the loads of each phase expressed by spatial features, successfully simplifying the expression form of the spatio-temporal characteristics of the three-phase load of the distribution transformer and enhancing the effectiveness of feature expression, ultimately supporting the efficient learning of the weekly pre-split phase load prediction model of the distribution transformer.

[0185] Taking the three-phase load IMIC Taking the scenario where both (X; Y) values are greater than the threshold as an example, at this time, the feature combination includes the three phases A, B, C and the overall historical load data, and the output target is the weekly-ahead predicted load of phases A, B and C.

[0186] Each frame of feature map is encoded in the form of a three-dimensional tensor, with a total of four channels. Each channel image contains N 2 pixel points. Pixel point K (i,j,W) corresponds to X load sampling points arranged in time series within a given period. Depending on the given period and time granularity, X can take values such as 168 (time granularity of 1 hour within 1 week), 1344 (time granularity of 15 minutes within 2 weeks), etc. W is the channel, representing phases A, B, C and the overall phase D respectively. K is the load rate, and the calculation formula is (4); pixel point sign is the flag bit, with a total number of Y. sign = 1 indicates that the phase represented by this channel is the prediction target, and sign = 0 indicates that the phase represented by this channel is a reference feature and is not used as a prediction target. The calculation relationship among N, X and Y is expressed as formula (5):

[0187] K j = I j / (I N ) j , j ∈ {A, B, C, D} (4)

[0188] In the formula, I is the phase current, and I N is the rated phase current.

[0189]

[0190] Through the spatial characterization technology of the inter-phase load coupling relationship and the multi-periodic characteristics of the load, the weekly-ahead phase-separated load prediction task of the distribution transformer is transformed into a spatio-temporal prediction task for multi-channel feature maps. At the same time, requirements are put forward for the spatial feature mining ability and the temporal feature extraction ability of the model.

[0191] 2. A weekly-ahead phase-separated load prediction model based on SwinLSTM-D is proposed to fully mine the spatio-temporal characteristics of the three-phase load of the distribution transformer;

[0192] 3. To meet the high-precision prediction requirements of the early warning task under light load and heavy overload conditions, a dual weighted hybrid loss function (DWH) is constructed to optimize the attention allocation mode of the model. The main steps are as follows:

[0193] Considering introducing the non-linear sensitivity characteristics of MAPE to construct the main part of the DWH loss function to correctly guide the model training direction, specifically as follows:

[0194] 1) Enhancement of heavy and overload conditions: For actual load values far from 0, by constructing an adaptive weight term as a function to introduce non-linear sensitivity characteristics, it generates significant losses during inaccurate predictions under heavy and overload conditions, and the higher the degree of heavy and overload, the greater the potential loss risk, forcing the model to pay attention to the prediction accuracy under heavy and overload conditions.

[0195] 2) Robustness of intermediate values: For actual load values within the normal range, the original MAPE calculation method is adopted, and through moderate losses, the stable performance of the model in such common scenarios is ensured, and unnecessary attention dispersion is avoided.

[0196] In addition, for light load conditions, this paper constructs a weighted mean square error term. Based on the characteristics of its error square, only significant penalties are given to large deviations under light load conditions to reasonably attract the attention of the model and optimize the model's attention to light load conditions. Finally, the calculation formula of the DWH loss function can be expressed as:

[0197]

[0198] In the formula: N represents the total number of samples, A i is the actual load value, F i is the predicted value corresponding to A i , ω is the weight, which can determine the attention degree of the model to light load conditions, τ 1 and τ 2 are the function segmentation thresholds, responsible for defining the load ranges of light load and heavy and overload conditions, and both ω and τ 1 and τ 2 are hyperparameters.

[0199] To sum up, the DWH loss function is based on a segmented weighted strategy, introducing non-linear sensitivity characteristics, ensuring that the model faces significant loss risks under heavy and overload conditions, and reasonably paying attention to large prediction deviations under light load conditions, thereby forcing the focus of the model's attention, automatically adjusting the evolution direction, and ensuring the performance of the prediction model under light load and heavy and overload conditions.

[0200] 4. Refer to Figure 2 , and a method for depicting the dynamic safety load domain of distribution transformers considering the constraints of the hot spot temperature rise of distribution transformers and accurately evaluating the ultimate load capacity of distribution transformers is proposed. The main steps are as follows:

[0201] 1) Adopt a calculation model for the hot spot temperature of distribution transformers considering three-phase unbalanced conditions to solve the hot spot temperature and top oil temperature in the next week, as follows:

[0202] The calculation model for the hot spot temperature of distribution transformers under three-phase unbalanced conditions constructed based on the thermoelectric analogy theory, as shown in Equation (7), is an accurate and efficient solution for calculating the hot spot temperature of distribution transformers.

[0203]

[0204] Where: θ amb is the ambient temperature, R oil is the heat transfer thermal resistance between the transformer oil and the external environment, R a , R b , R c are respectively the heat transfer thermal resistances between the three-phase windings and the transformer oil, C oil , C a , C b , C c , C fe are respectively the heat capacities of the transformer oil, the three-phase windings and the iron core, q fe is the iron core loss, θ hs,a , θ hs,b , θ hs,c , θ fe , θ to are respectively the hot spot temperatures and the top oil temperatures of the A, B, C phase windings and the iron core, q cu,a , q cu,b , q cu,c are respectively the load losses of the three-phase windings, which are determined by the load factor, and the calculation formula is as follows:

[0205] q cu,i =(K i (I N ) i ) 2 R i , i ∈ {A, B, C} (8)

[0206] Where: K i is the phase load factor, (I N ) i is the phase rated current, R i is the sum of the high-voltage winding resistance reduced to the low-voltage side and the low-voltage winding resistance.

[0207] Based on Equations (7) and (8), the hot spot temperatures and the top oil temperatures of the A, B, C phase windings and the iron core of the distribution transformer considering the three-phase unbalanced operating conditions under the given external conditions can be accurately captured.

[0208] 2) Based on the weekly prediction data of the hot spot temperature and the top oil temperature, characterize the dynamic safety load domain of the distribution transformer considering the hot spot temperature rise constraint of the distribution transformer, and accurately evaluate the ultimate load capacity of the distribution transformer, as follows:

[0209] Define the dynamic safety load domain of the distribution transformer as follows: Considering the three-phase unbalanced operating conditions, at any given moment, the limit load space that ensures the distribution transformer meets the hot-spot temperature rise constraint within the specified safe operating time limit. Given that different initial temperature distributions inside the distribution transformer at different given moments will result in different hot-spot temperature rises, the safety load domain of the distribution transformer is dynamic and needs to be characterized separately according to the temperature distribution inside the distribution transformer at different given moments. The steps are as follows:

[0210] ①. Acquisition of distribution transformer samples: For the target distribution transformer, comprehensively collect the key thermal and electrical parameters required by the hot-spot temperature calculation model, and obtain the pre-week phase-separated load prediction results for the target warning week.

[0211] ②. Construction of the thermal circuit model: Establish the thermal circuit model of the distribution transformer based on the thermoelectric analogy theory.

[0212] ③. Calculation of the temperature curve: Based on the pre-week phase-separated load prediction results of the distribution transformer, continuously solve the thermal circuit model by dynamically adjusting the initial temperature to accurately calculate the pre-week temperature change curves of the hot-spot temperatures of the three-phase windings and the iron core and the top oil temperature.

[0213] ④. Setting of operating limits: Follow the design specifications and safety standards of the distribution transformer to stipulate the hot-spot temperature rise constraint and the safe operating time limit.

[0214] ⑤. Construction of the dynamic domain: For each moment of the pre-week temperature prediction curve, simulate the temperature change process of the distribution transformer within the safe operating time limit under different load scenarios one by one, identify and define the limit load space of the distribution transformer that meets the hot-spot temperature rise constraint, and finally complete the characterization of the dynamic safety load domain of the distribution transformer.

[0215] 5. Refer to Figure 3 , and propose a pre-week phase-separated overload and heavy-load warning strategy for the distribution transformer based on the dynamic safety load domain of the distribution transformer. The main steps are as follows:

[0216] 1) Since all load points within the dynamic safety load domain of the distribution transformer meet the hot-spot temperature rise constraint within the safe operating time limit, therefore, set the domain boundary as the overload warning boundary. When the load point approaches this boundary, it means that the distribution transformer is about to enter the overload condition.

[0217] 2) Set 80% of the overload warning boundary as the heavy-load warning boundary. When the load point reaches this boundary, it indicates that the distribution transformer has entered the heavy-load condition.

[0218] 3) For each moment in the pre-week phase-separated load prediction results of the distribution transformer, put the predicted load point into the domain space. By evaluating the spatial relative positions of the load point with the heavy-load warning boundary and the overload warning boundary, the pre-week phase-separated overload and heavy-load warning results of the distribution transformer at this moment can be finally determined.

[0219] Example 13:

[0220] A method for verifying the weekly pre - phase - separated heavy - overload warning of distribution transformers considering the spatio - temporal characteristics of three - phase loads and the dynamic safety load domain of distribution transformers, the main steps are as follows:

[0221] 1) Use the measured data of distribution transformers from a power supply bureau in southern China from December 1, 2022 to November 30, 2023 for example verification, and the sampling frequency of load data is 1 hour;

[0222] 2) Explore the spatio - temporal characteristics of the three - phase loads of distribution transformers, and strengthen the spatio - temporal feature expression of the three - phase loads of distribution transformers based on the dynamic selection technology of feature combination and the spatial representation technology;

[0223] 3) Based on SwinLSTM - D, establish a weekly pre - phase - separated load prediction model for distribution transformers to fully explore the spatio - temporal characteristics of the three - phase loads of distribution transformers;

[0224] 4) Facing the high - precision prediction requirements in light - load and heavy - overload conditions for the warning task, construct a dual - weighted hybrid loss function (Dual Weighted Hybrid Loss, DWH) to optimize the attention distribution mode of the model;

[0225] 5) Characterize the dynamic safety load domain of distribution transformers considering the temperature rise constraints of hot spots of distribution transformers to accurately evaluate the ultimate load - carrying capacity of distribution transformers;

[0226] 6) Based on the dynamic safety load domain of distribution transformers, follow the proposed strategy to conduct weekly pre - phase - separated heavy - overload warning for distribution transformers;

[0227] 7) To verify the performance of the proposed prediction method, this paper sets a total of 5 methods for comparison.

[0228] Method 1: Only consider the load time series, including BiLSTM and GRU.

[0229] Method 2: Consider both the load time series and multi - periodicity, including MDGRU - Seq2Seq.

[0230] Method 3: Consider both the load time series and the inter - phase load coupling relationship, including CNN - LSTM.

[0231] Method 4: Consider the load time series, multi - periodicity and inter - phase load coupling relationship, including MS - STGNN.

[0232] Method 5: The method in this paper.

[0233] The overall accuracy obtained is shown in Table 1, and the accuracy in light - load and heavy - overload conditions in the test set is shown in Table 2:

[0234] Table 1 Annual accuracy test results

[0235]

[0236] Table 2 Precision test results under light load and heavy overload conditions throughout the year

[0237]

[0238] It is assumed that in different distribution transformer datasets, the model with the best indicators in the control algorithm is used as the benchmark model.

[0239] First of all, the method proposed in this paper shows the best prediction performance. In Table 2, the RMSE errors of the annual precision measured by the method in this paper are 81.25% and 80.95% of the benchmark model respectively, and the MAPE errors are 80.14% and 72.58% of the benchmark model respectively. Among the 5 comparison models, MS-STGNN achieved the highest prediction accuracy, MDGRU-Seq2Seq and CNN-LSTM performed similarly and tied for the second place, and GRU and BiLSTM had the lowest accuracy. The above results show that: 1) Compared with GRU and BiLSTM that only consider the load time series, capturing the coupling relationship between phase loads and the multi-periodic characteristics of loads can significantly improve the prediction accuracy. 2) Compared with MS-STGNN, the method in this paper has significant advantages in the tasks of mining the coupling relationship between phase loads and extracting the multi-periodic characteristics of loads by virtue of the feature engineering that conforms to the spatio-temporal characteristics of three-phase loads and the SwinLSTM-D model guided by DWH.

[0240] Secondly, the method in this paper has the best adaptability under light load and heavy overload conditions. The proposed model in this paper obtained the best precision test results under light load and heavy overload conditions. The RMSE errors are 58.68% and 78.99% of the benchmark model respectively, and the MAPE errors are 67.53% and 69.74% of the benchmark model respectively. Under light load and heavy overload conditions, the prediction accuracy of the benchmark model has a certain degree of attenuation, while the prediction accuracy of the proposed model in this paper has increased instead. The above results show that, relying on the DWH loss function facing the requirements of the early warning task, the proposed model in this paper fully adapts to light load and heavy overload conditions and has significant advantages compared with the comparison models.

[0241] In summary, the following conclusions can be drawn from this group of experiments: The proposed prediction method in this paper has excellent prediction performance and meets the high-precision prediction requirements of the early warning task for light load and heavy overload conditions of each phase, and can support the accurate early warning of the distribution transformer heavy overload problem.

[0242] 8) To verify the performance of the proposed early warning method, this paper compares it with the current method (using the proportion of the load in the static rated capacity as the early warning standard).

[0243] By depicting the dynamic safety load domain of the distribution transformer and conducting a heavy overload assessment based on the depiction results, it can be seen that the early warning situation during this period is shown in Table 4. Among them, the heavy overload assessment results at t = 11, t = 39, and t = 42 are different from those of the current method, as shown in Table 5.

[0244] Table 4 Transformer Summer Distribution Transformer Heavy Overload Early Warning Results

[0245]

[0246] Table 5 Heavy Overload Evaluation Results at t = 11, t = 39, and t = 42

[0247]

[0248] The following conclusions can be drawn from this group of experiments: 1) As can be seen from Table 5, when considering the temperature rise constraint of the distribution transformer hot spot, there are differences between the load capacity of the distribution transformer under the three-phase unbalanced condition and the rated capacity, and the traditional method cannot give accurate early warning results. 2) As can be seen from Table 4 and Table 5, based on the dynamic safety load domain of the distribution transformer described in this paper, the ultimate load capacity of the distribution transformer under the temperature rise constraint of the hot spot can be accurately evaluated, and finally, the precise early warning of the heavy overload of the distribution transformer considering the three-phase unbalanced condition can be realized.

Claims

1. A phase-by-phase heavy overload warning method for distribution transformers before the week considering the spatiotemporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers, characterized in that: The following steps are involved: 1) Based on the spatiotemporal characteristics of the three-phase load of the distribution transformer, a multi-channel characteristic diagram is constructed to characterize the inter-phase load coupling relationship and the multi-periodic characteristics of the load. 2) Establish a phase-by-phase load forecasting model for distribution transformers based on SwinLSTM-D; 3) Construct a double-weighted hybrid loss function to optimize the attention allocation mode of the phase-by-phase load forecasting model for distribution transformers before the week; 4) Characterize the dynamic safety load domain of the distribution transformer considering the hot spot temperature rise constraint of the distribution transformer to evaluate the ultimate load capacity of the distribution transformer; 5) Solve the phase-by-phase load prediction model for the distribution transformer before the week to obtain the load prediction results; compare the load prediction results with the dynamic safety load domain of the distribution transformer, and generate the phase-by-phase heavy overload warning information for the distribution transformer before the week.

2. According to claim 1, a phase-by-phase heavy overload warning method for distribution transformer cycle before considering the spatiotemporal characteristics of three-phase load and the dynamic safety load domain of distribution transformer is characterized in that: In step 1), the step of constructing a multi-channel characteristic diagram to characterize the phase-to-phase load coupling relationship and the load multi-periodicity characteristics includes: 1.1) Based on the X-phase load data of the distribution transformer X=(x 1 ,x 2 ,…,x n ) and Y phase load data Y=(y 1 ,y 2 ,…,y n ), construct the data set D = {(x i ,y i ),i=1,2,…,n}; n is the sequence length; Among them, the mutual information number of X-phase load data and Y-phase load data is as follows: Where: I MI (X; Y) is the mutual information coefficient between phase X and phase Y; p(x, y) is the joint probability density function of phase X and phase Y; p(x) and p(y) are the marginal probability density functions of phase X and phase Y, respectively; 1.2) Discretize the data set D on a two-dimensional plane and divide it into a×b grids, denoted as grid G, and calculate the maximum mutual information I*MI(D| G ,a,b); The maximum mutual information I*MI(D| G ,a,b) are as follows: Where: D| G is the probability distribution of a point in D falling into G; I MI (D∣ G ,a,b) is the probability distribution D| G The estimated mutual information between phase X and phase Y; 1.3) Calculate the maximum information coefficient I MIC (X; Y), that is: Where: B(n) is the constraint function that determines the maximum number of grids; 1.4) Based on the maximum information coefficient I MIC (X; Y), quantify the coupling degree between loads in different seasons; 1.5) For each phase load, select the historical load data and I MIC Other phase historical load data (X; Y) greater than a preset threshold are used as input features; 1.6) For each phase in the input feature, the single-phase time series load sequence is encoded into a single-channel two-dimensional feature map, so that the prediction network processes the load information from line to surface; 1.7) The load data of each phase in the input feature is independently constructed into adjacent channels, and the features between the channels are interactively fused to form a spatial correlation pattern; 1.8) Repeat step 1.7) to encode the input characteristic load data into a multi-channel feature map, and each phase load is independently constructed as an adjacent channel; wherein the inter-phase load coupling relationship is reflected as the spatial dependence relationship of the corresponding areas of the adjacent channels, and each channel contains the multi-periodic characteristics of each phase load expressed by spatial features.

3. According to claim 2, a phase-by-phase heavy overload warning method for distribution transformer cycle before considering the spatiotemporal characteristics of three-phase load and the dynamic safety load domain of distribution transformer is characterized in that: Maximum information coefficient I MIC (X; Y) is used to quantify the degree of coupling between the three phases and the overall load; I MIC (X; Y) range is [0,1], when the I MIC The larger the (X;Y) value, the deeper the coupling.

4. According to claim 1, a phase-by-phase heavy overload warning method for distribution transformer cycle before considering the spatiotemporal characteristics of three-phase load and the dynamic safety load domain of distribution transformer is characterized in that: The input data of the pre-week phase load forecasting model of distribution transformer based on SwinLSTM-D is the multi-channel feature map, and the output data is the phase load forecast value.

5. The method for early warning of heavy overload of distribution transformers before a cycle considering the spatiotemporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers according to claim 1 is characterized in that: In step 3), the double weighted hybrid loss function DWH is as follows: Where: N represents the total number of samples, A i is the actual load value, F i For A i The corresponding predicted value, ω is the weight, τ1 and τ2 are the function segmentation thresholds, and ω, τ1 and τ2 are all hyperparameters.

6. The method for early warning of heavy overload of distribution transformers before a cycle considering the spatiotemporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers according to claim 1 is characterized in that: In step 4), the steps of describing the dynamic safety load domain of the distribution transformer considering the hot spot temperature rise constraint of the distribution transformer include: 4.1) Using the distribution transformer hot spot temperature calculation model taking into account the three-phase unbalanced working condition, solve the A, B, C three-phase winding and core hot spot temperature and top oil temperature of the distribution transformer in the next week; 4.2) Based on the week-ahead prediction data of hot spot temperature and top oil temperature, the dynamic safety load domain of the distribution transformer considering the hot spot temperature rise constraint of the distribution transformer is characterized to evaluate the ultimate load capacity of the distribution transformer.

7. A phase-by-phase heavy overload warning method for distribution transformer cycle before the cycle considering the spatiotemporal characteristics of three-phase load and the dynamic safety load domain of distribution transformer according to claim 6, characterized in that: The distribution transformer hot spot temperature calculation model is as follows: Where: θ amb is the ambient temperature, R oil is the heat transfer resistance between transformer oil and external environment, R a , R b , R c are the heat transfer resistance of the three-phase winding and transformer oil, C oil , C a , C b , C c , C fe are the heat capacities of transformer oil, three-phase winding and iron core, respectively, and q fe is the core loss, θ hs,a ,θ hs,b ,θ hs,c ,θ fe ,θ to are the hot spot temperature and top oil temperature of the three-phase windings A, B, and C as well as the core, q cu,a ,q cu,b ,q cu,c are the load losses of the three-phase windings respectively; Among them, the load loss q cu,i As shown below: q cu,i =(K i (I N ) i ) 2 R i ,i∈{a,b,c}(7) Where: K i is the phase load factor, (I N ) i is the phase rated current, R i It is the sum of the high voltage winding resistance converted to the low voltage side and the low voltage winding resistance.

8. The method for early warning of heavy overload of distribution transformers before a cycle considering the spatiotemporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers according to claim 6 is characterized in that: In step 4.2), the steps of describing the dynamic safety load domain of the distribution transformer considering the hot spot temperature rise constraint of the distribution transformer include: 4.2.1) Collect the thermal and electrical parameters of the target distribution transformer and obtain the phase load forecast results one week in advance; 4.2.2) Establish a distribution transformer hot spot temperature calculation model based on thermoelectric analogy theory; 4.2.3) Based on the phase load prediction results of the distribution transformer before the week, the thermal circuit model is continuously solved by dynamically adjusting the starting temperature to calculate the temperature change curves of the hot spot temperature of the three-phase winding and the core and the top oil temperature before the week; 4.2.4) Comply with distribution transformer design specifications and safety standards, and specify hot spot temperature rise constraints and safe operation limits. 4.2.5) For each moment of the week-ahead temperature prediction curve, simulate the temperature change process of the distribution transformer within the safe operation limit under different load scenarios one by one, identify and define the distribution transformer limit load space that meets the hotspot temperature rise constraint, and complete the characterization of the dynamic safe load domain of the distribution transformer.

9. The method for early warning of heavy overload of distribution transformers before a cycle considering the spatiotemporal characteristics of three-phase loads and the dynamic safety load domain of distribution transformers according to claim 1 is characterized in that: In step 5), before generating the pre-cycle phase-by-phase heavy overload warning information of the distribution transformer, the dynamic safety load domain boundary of the distribution transformer is set as the overload warning boundary, and 80% of the overload warning boundary is the heavy load warning boundary.

10. A phase-by-phase heavy overload warning method for distribution transformer cycle before the cycle considering the spatiotemporal characteristics of three-phase load and the dynamic safety load domain of distribution transformer according to claim 9, characterized in that: If the load prediction result reaches the overload warning boundary, the generated phase-by-phase heavy overload warning information before the distribution transformer cycle is an overload condition warning; If the load forecast result reaches the heavy load warning boundary, the generated pre-cycle phase heavy overload warning information of the distribution transformer is a heavy load condition warning.

Images