Transformer oil temperature prediction method based on state visible diagram

Abstract

The invention discloses a transformer oil temperature prediction method based on a state visible graph, and the method comprises the steps: obtaining an initial time-state sequence of transformer oil temperature based on a time sequence, the initial time-state sequence comprises a plurality of elements, and the elements comprise time data and node state data; performing coarse graining and local fluctuation characteristic analysis on the initial time-state sequence of the oil temperature of the transformer in sequence to obtain a state sequence of the time sequence; constructing a state visible graph based on the state sequence of the time sequence; and acquiring the node state at the current moment, and predicting the oil temperature value of the transformer at the next moment based on the state visible graph and the node state at the current moment. According to the method, a more intuitive and more accurate transformer oil temperature prediction result can be obtained, the power distribution problem is effectively optimized, and unnecessary waste is reduced.

Description

Technical Field

[0001] The invention relates to the technical field of time series prediction of power networks, and in particular to a transformer oil temperature prediction method based on a state visible graph. Background Art

[0002] The power distribution problem is that the power grid manages the distribution of power to different user areas according to the sequentially changing demand. However, it is difficult to predict the future demand of a specific user area because it varies with different factors such as weekdays, holidays, seasons, weather, temperature, etc. Existing prediction methods cannot be applied to high-precision long-term predictions of long-term real-world data, and any wrong predictions may have serious consequences. Therefore, there is currently no effective way to predict future power consumption, and managers have to make decisions based on empirical values, and the threshold of empirical values ​​is usually much higher than the actual demand. Conservative strategies lead to unnecessary waste of electricity and equipment depreciation. It is worth noting that the oil temperature of the transformer can effectively reflect the working condition of the power transformer. Therefore, predicting the oil temperature of the transformer can also try to avoid unnecessary waste.

[0003] Time series prediction is a modeling method for a certain phenomenon or process in the real world. Time series prediction takes historical data as the research object, models the observed time series, analyzes the potential dependencies in historical data, obtains its changing rules in the modeling process, and predicts the future by understanding the past. Therefore, we can perfectly combine the time prediction series algorithm and apply it to predict the oil temperature of the transformer to predict future electricity consumption and avoid unnecessary waste.

[0004] The oil temperature network of real transformers is full of uncertainty. The main reason is that the interaction between different objects in the system is difficult to model in an accurate way, and the knowledge of the system is not complete. Usually, these uncertainties and fluctuations will bring prediction errors to the prediction model. Therefore, the inventors have conducted in-depth research on uncertainty and proposed many methods to deal with uncertainty in time series prediction, such as grayscale analysis and fuzzy set theory. In recent years, complex networks have attracted widespread attention because they can effectively model the complexity of complex systems in the real world, especially the interaction between different nodes or objects. At the same time, some studies have shown that complex networks can assist time series prediction, and the study of complex networks can be used to explore the effective information hidden in time series. Time series analysis based on complex networks has become a unique perspective for time series prediction. The inventors believe that time series are closely related to complex networks, and have conducted many studies on the application of complex network theory in time series analysis, resulting in many new algorithms for mapping time series to complex networks. For example, visibility graph (VG), weighted visibility graph, pseudo-periodic time series transformation algorithm, phase space reconstruction algorithm and phase space coarse-grained algorithm, etc. Therefore, modeling the oil temperature data of the transformer as a complex network for time series prediction will produce more valuable results. The following are some studies on time series analysis using complex network technology.

[0005] The prediction model based on the visible graph first uses the link prediction technology to obtain the similar nodes of the target node, and then predicts the value of the target node based on the correlation between the similar nodes and the target node. The prediction model based on the visible graph generally uses the trend extension method to make predictions, that is, it is assumed that the trend of the node to be predicted is consistent with the trend of the target node. Such a prediction method is easily affected by local maximum and minimum values ​​when encountering a periodic time series with frequent fluctuations, resulting in an increase in the prediction error of the prediction method.

[0006] At the same time, in the phase space coarse-grained algorithm model based on the phase space reconstruction algorithm, the scalar time series is embedded into the high-dimensional symbol sequence in the phase space through the phase space reconstruction theory and the coarse-grained algorithm, which can effectively characterize the dynamic characteristics of the time series. However, the phase space coarse-grained algorithm will lose some accuracy when constructing the symbol sequence through the fluctuation sequence of the time series, and is easily affected by the adverse effects of parameter selection when dividing the fluctuation mode sequence. On the other hand, the visible graph has the characteristics of connectivity, non-directivity and invariance under affine transformation, which can effectively describe the correlation between nodes in the transformer oil temperature prediction, but lacks the description of the fluctuation characteristics of the time series. The DFN model based on the phase space coarse-grained algorithm can effectively capture the fluctuation characteristics of the time series, but due to the influence of the model parameter selection, it is easy to lose the prediction accuracy during the retrospective prediction. Therefore, the present invention proposes a state visible graph algorithm based on the visible graph and the phase space coarse-grained algorithm to realize the oil temperature prediction of the transformer. Summary of the invention

[0007] To solve the above technical problems, the present invention constructs a state visible graph based on the visible graph and the phase space coarse-grained algorithm, and proposes a time series prediction algorithm based on the state visible graph. This technology combines the advantages of both, and can obtain more intuitive and accurate transformer oil temperature prediction results, effectively optimize power distribution problems, and reduce unnecessary waste.

[0008] The present invention provides a transformer oil temperature prediction method based on a state visibility graph, comprising the following steps:

[0009] Acquire an initial time-state sequence of transformer oil temperature based on the time series, wherein the initial time-state sequence includes several elements, including: time data and node state data;

[0010] The initial time-state sequence of the transformer oil temperature is subjected to coarse-graining and local fluctuation characteristic analysis in sequence to obtain a state sequence of the time series;

[0011] constructing a state visibility graph based on the state sequence of the time series;

[0012] The node status at the current moment is obtained, and the oil temperature value of the transformer at the next moment is predicted based on the state visibility graph and the node status at the current moment.

[0013] Optionally, the process of processing and transforming the initial time-state sequence to construct a state visibility graph includes:

[0014] Calculating the first-order mean absolute difference of the initial time-state sequence to obtain a first-order mean absolute difference value;

[0015] defining a plurality of symbols based on the first-order mean absolute difference value;

[0016] After the elements in the initial time-state sequence are coarse-grained, they are combined with the symbols to construct a symbol sequence based on the element mapping relationship;

[0017] Converting the symbol sequence into an oil temperature state sequence;

[0018] The state visible graph is obtained by converting and calculating the oil temperature state sequence based on a sliding visible graph algorithm.

[0019] Optionally, the calculation formula for calculating the first-order mean absolute difference value is:

[0020]

[0021] In the formula, R average represents the first-order mean absolute difference, y i Represents the transformer oil temperature value of the i-th node.

[0022] Optionally, a calculation formula for defining a number of symbols based on the first-order mean absolute difference value is:

[0023]

[0024] In the formula, k represents the total number of matches, y min Represents the minimum node state data in the initial time-state sequence, y max Represents the maximum node state data in the initial time-state sequence.

[0025] Optionally, the element mapping relationship is:

[0026] Yes min +(m-1)R average ≤y t

[0027] And y t <y min +mR average

[0028] In the formula, γ m Represents the mth symbol.

[0029] Optionally, the local fluctuation characteristics include: a steady state, a rising state, a falling state, a maximum state and a minimum state.

[0030] Optionally, the process of predicting the oil temperature value of the transformer at the next moment based on the state visibility graph and the node state at the current moment includes:

[0031] Calculate the state node of the node state at the current moment of observation in the oil temperature state sequence;

[0032] Obtaining an in-degree node set and an out-degree node set of the state node in the state visible graph;

[0033] Performing weighted summation on the out-degree node set to obtain an out-degree prediction value;

[0034] The in-degree node set is calculated using a trend extension method to obtain an in-degree prediction value;

[0035] The out-degree prediction value and the in-degree prediction value are weightedly summed to obtain the transformer oil temperature prediction value at the next moment.

[0036] Optionally, the calculation formula of the transformer oil temperature prediction value at the next moment is:

[0037]

[0038] In the formula, Indicates the predicted value of transformer oil temperature at the next moment, represents the predicted in-degree value, Indicates the predicted value of degree,

[0039] The present invention has the following technical effects:

[0040] The present invention constructs a state visible graph based on the visible graph and the phase space coarse-grained algorithm, and proposes a time series prediction algorithm based on the state visible graph. This technology combines the advantages of both, and can obtain more intuitive and accurate transformer oil temperature prediction results, effectively optimize power distribution problems, and reduce unnecessary waste. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0042] Figure 1 It is a visualization of the data set in the embodiment of the present invention, wherein Figure a represents the ETTh1 data set, Figure b represents the PEMS-07 data set, Figure c represents the PEMS-BAY data set, Figure d represents the METR-LA data set, Figure e represents the SOLAR data set, and Figure f represents the COVID-19 data set;

[0043] Figure 2The state diagrams mapped to the time series with cycle numbers of 100, 200 and 500 in the embodiment of the present invention are visible, wherein Figure a represents the cycle number 100, Figure b represents the cycle number 200, and Figure c represents the cycle number 500.

[0044] Figure 3 is the degree distribution of the random sequence state visible graph in the embodiment of the present invention, wherein graph a represents a uniform random sequence graph, and graph b represents a Gaussian random sequence graph;

[0045] Figure 4 The random sequence state in the embodiment of the present invention can be seen as the degree distribution of the graph, wherein graph a represents the sequence graph and graph b represents the degree distribution graph;

[0046] Figure 5 It is an implementation flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Embodiment 1

[0049] The core of this technology is to combine the characteristics of visible graph and phase space coarse-grained algorithm in the construction process of transformer oil temperature network through state visible graph algorithm, coarse-grained symbolization of time series data, and conversion of time series nodes into state nodes of state space, which have value attributes and fluctuation attributes of time series. At the same time, visibility is used to determine whether there are edges between state nodes in the network. The transformer oil temperature state visible graph not only describes the local dynamic characteristics of the nodes, but also pays attention to the correlation between non-adjacent nodes, so it can predict the accuracy of transformer oil temperature changes.

[0050] like Figure 5 As shown, this embodiment discloses a transformer oil temperature prediction method based on a state visibility graph, and the specific implementation process includes:

[0051] The initial time-state sequence of the transformer oil temperature is obtained based on the time series. The initial time-state sequence includes several elements, including: time data and node state data, specifically including:

[0052] 1. Construction process of the state visibility graph of the transformer oil temperature network

[0053] Given a transformer oil temperature time series X = {(t1,y1),(t2,y2),...,(tn ,y n )}, the minimum value of sequence X is y min , the maximum value is y max ,The construction process of the transformer oil temperature state diagram is as follows.

[0054] Step 1: Calculate the first-order mean absolute difference R of the transformer oil temperature time series X average , the formula is as follows:

[0055]

[0056] The initial time-state sequence of transformer oil temperature is subjected to coarse-graining and local fluctuation characteristic analysis to obtain the state sequence of the time series, including:

[0057] Step 2: Coarse-grain the transformer oil temperature time series X to obtain the symbol sequence S0. Define k symbols {γ0,γ1,...,γ k-1},in

[0058] The value of each element in the transformer oil temperature time series X is coarse-grained into an element in the symbol sequence S0. The mapping relationship between the sequence X and the elements in S0 is as follows:

[0059]

[0060] Step 3: Calculate the transformer oil temperature network node (x i ,y i ) local fluctuation characteristics, converting the symbol sequence S0 into a state sequence S. This paper takes five fluctuation states of the node to define the local fluctuation characteristics of the transformer oil temperature time series {δ0,δ1,δ2,δ3,δ4}, which are steady state, rising state, falling state, maximum state and minimum state respectively. Combining the value state and fluctuation characteristics of the node, the state sequence S of the transformer oil temperature time series is defined as follows:

[0061] S={s kd |s kd =(γ k ,δ d ),1≤k≤n,0≤d≤4} (3)

[0062] The calculation formula of d is as follows:

[0063]

[0064] When d=0, it means the current node is in a stable state, when d=1, it means the current node is in an ascending state, when d=2, it means the current node is in a descending state, when d=3, it means the current node is in a minimum state, when d=4, it means the current node is in a maximum state, and the trend extension method is used to evaluate the fluctuation state of the node for the first and last nodes of the sequence. Each node in the time series uniquely corresponds to a state of the state sequence, and each state node corresponds to several time series nodes.

[0065] The state visibility graph is constructed based on the state sequence of the time series, including:

[0066] Step 4: Based on the transformer oil temperature state sequence S obtained in step 3, the time axis direction is used as the visibility direction, the window length is defined as L, the sliding step is defined as 1, and the sliding visibility graph algorithm is combined to construct a directed and weighted state visibility graph (SVG) of the transformer oil temperature network. The specific steps are shown in Algorithm 1.

[0067] Algorithm 1 constructs an algorithm based on the state visibility graph of the sliding window.

[0068] Input: Transformer oil temperature time series X = {(t1,y1),(t2,y2),...,(t n ,y n )}, sliding window length L, state sequence S = {(s k ,s d )|1≤k≤n,0≤d≤4}.

[0069] Output: The status of the transformer oil temperature network can be seen in Figure SVG.

[0070] The specific implementation process includes:

[0071] Initialize the visible graph VG with X as the node set, and initialize the state visible graph SVG with S as the node set

[0072] Divide the transformer oil temperature time series X into two subsequences

[0073] X0={(t1,y1),(t2,y2),...,(t L ,y L )}

[0074] X1={(t L+1 ,y L+1 ),(t L+2 ,y L+2 ),...,(t n ,y n )}

[0075] Get the index sequence Index of sequence X0 from large to small

[0076] FOR i←1to L DO

[0077] Get the node with the maximum value of the current visibility that needs to be calculated from the sequence Index, cur←Index[i]

[0078] Get all connected nodes of cur from VG, connected_nodes←G[cur]

[0079]

[0080] Determine the visibility of node cur in the sequence X0[left,right] and obtain the visible node list List1

[0081] Update VG with the nodes in node List1 and cur, and get the corresponding state node from S to update SVG. If the edge in SVG already exists, the weight is increased by 1.

[0082] END FOR

[0083] FOR node in X1 DO

[0084] Remove the first node of the transformer oil temperature sequence X0 and add node to the end of the sequence X0

[0085] Determine the visibility of node node in the transformer oil temperature sequence X0 and obtain the visible node list List2

[0086] Update SVG using the nodes in node List2 and the state nodes corresponding to node. If the edge already exists in SVG, the weight w of the edge is increased by 1.

[0087] END FOR

[0088] RETURN SVG

[0089] Get the node status at the current moment, and predict the oil temperature value of the transformer at the next moment based on the state visibility graph and the node status at the current moment, including:

[0090] 2. Transformer oil temperature network prediction algorithm based on state visibility graph

[0091] Since the nodes in the transformer oil temperature network (t n ,y n ) is the observation node closest to the prediction node in the oil temperature change time series X (t n+1 ,y n+1 ), which has the most direct impact on the predicted value of transformer oil temperature. Therefore, this paper chooses to use node (tn ,y n ) is the basis for t n+1 The oil temperature value of the transformer at the moment is predicted. The prediction steps are as follows:

[0092] Calculate the oil temperature change network nodes (t n ,y n ) in the state node s in the oil temperature state sequence S n .

[0093] Get n The in-degree node set S in the state visibility graph SVG in and the out-degree node set S out .

[0094] For the set S out , perform weighted summation on the out-degree node set to get the predicted value The calculation formula is as follows:

[0095]

[0096] where w ni Indicates the state visible graph SVG from node s n To node s i The weight of the edge; |s i | represents node s i In the SVG construction process, this paper records the sequence elements converted into state nodes and calculates the mathematical expectation of the state nodes by averaging.

[0097] For the set S in , calculate the predicted value by the trend outer edge method The calculation formula is as follows:

[0098]

[0099] Prediction of transformer oil temperature Perform weighted summation to obtain the transformer oil temperature prediction value As the ultimate goal. The calculation formula is as follows:

[0100]

[0101] in

[0102] In order to briefly explain the construction process and prediction algorithm of the transformer oil temperature change network visible graph, we use the sequence X = {1, 2, 3, 4, 5} as an example to construct the state visible graph and predict the value of the transformer oil temperature at the next moment. First, we calculate the first-order mean absolute difference of the sequence, R averge= 1, then define k = (5-1) / 1+1 = 5 symbols {γ0, γ1, γ2, γ3, γ4}, and then use formula (13) to coarse-grained symbolize the sequence X. For example, the coarse-grained symbol of 2 is γ1 = γ (2-1) / 1 , so the symbol sequence corresponding to X is S0 = {γ0,γ1,γ2,γ3,γ4,γ3,γ4}. Next, use formula (15) to calculate the local fluctuation characteristics of each element in the sequence X. The local fluctuation characteristics of the first element 1 and the last element 5 are in the rising state δ1 according to the trend outer edge method, so the fluctuation characteristic sequence of the sequence X is X1 = {γ0,γ1,γ2,γ3,γ4,γ3,γ4}. Combining the sequences X0 and X1, there is a state sequence S = {(γ0,δ1),(γ1,δ1),(γ2,δ1),(γ3,δ1),(γ4,δ4),(γ3,δ3),(γ4,δ1)}, in which each element is unique and corresponds to the seven state nodes in SVG.

[0103] After obtaining the state sequence of transformer oil temperature change, the elements in sequence X are used to determine the visibility between the corresponding state nodes. For example, the first two nodes (γ0, δ1) and node (γ1, δ1) of sequence S correspond to the first two elements 1 and 2 in sequence X, so there is an edge between them with weight w = 1. SVG edge set

[0104] E={[(γ0,δ1),(γ1,δ1),w=1],[(γ2,δ1),(γ3,δ1),w=1],

[0105] [(γ3,δ1),(γ4,δ4),w=1],[(γ4,δ4),(γ3,δ3),w=1],

[0106] [(γ3,δ3),(γ4,δ1),w=1]}.

[0107] After the SVG is constructed, the value of the next node y8 in X is predicted. It is known that the last element in X is 5, and the corresponding state node is s7 = (γ4, δ1), with one in-degree node s6 = (γ3, δ3), and no out-degree node. The predicted value of the transformer oil temperature is calculated according to formulas (17) and (18) as follows:

[0108]

[0109] In order to verify the effectiveness of the solution proposed in this article, this example further conducts experiments to prove that:

[0110] 1. Dataset

[0111] We used the ETTh1 dataset, which is a two-year data set collected by researchers from a county in China, to verify that our technology can implement a time series prediction algorithm based on state visibility graphs. We also used four additional real datasets from other fields to verify that our technology can also be promoted and applied in other fields.

[0112] In the experiment, 500 to 1000 nodes were uniformly extracted from each sequence by downsampling to obtain the data trend of the time series. Figure 1 shown.

[0113] The data is introduced as follows:

[0114] (1) ETTh1 dataset. ETTh1 is a dataset of oil temperature of power transformers. It is a two-year data set collected by researchers from a county in China. The sampling interval of the data is 1 hour.

[0115] (2) PEMS-07 dataset. PEMS is a highway performance measurement system in California, USA, which collects traffic datasets of major cities in California. The PEMS system is divided into 12 regions, and PEMS-07 is the dataset of the seventh region.

[0116] (3) PEMS-BAY dataset. PEMS-BAY is a dataset collected by the California Highway Performance Measurement System in the San Francisco Bay Area, covering six months from January 1, 2017 to May 31, 2017.

[0117] (4) METR-LA dataset. METR-LA is a traffic information dataset of Los Angeles freeways in the United States. The dataset contains data from March 1, 2012 to June 30, 2012, a total of four months.

[0118] (5) SOLAR dataset. SOLAR is a solar power dataset for 2006 compiled by the National Renewable Energy Laboratory of the United States. The Alabama region dataset used in this paper. The present invention performs statistical analysis on all datasets, and the statistical results are shown in Table 1.

[0119] Table 1

[0120]

[0121] 2. Evaluation indicators

[0122] In order to evaluate the performance of the algorithm proposed in this paper, five commonly used test indicators are used to evaluate the prediction accuracy of the model.

[0123] (1) Mean Absolute Error (MAE)

[0124]

[0125] (2) Mean Square Error (MSE)

[0126]

[0127] (3) Root Mean Square Error (RMSE)

[0128]

[0129] (4) Mean Absolute Percentage Error (MAPE)

[0130]

[0131] (5) Mean Square Prediction Error (MSPE)

[0132]

[0133] Among the five evaluation indicators, y t and are the true value and predicted value of the transformer oil temperature change at time t. The five evaluation indicators analyze the prediction results of the transformer oil temperature from the overall error of the sequence. The smaller the value, the more accurate the prediction result of the transformer oil temperature.

[0134] 3. Validation of the Transformer Oil Temperature Network Status Visible Graph Algorithm

[0135] The key issue of the time series complex network conversion algorithm is whether the conversion graph can inherit some structural characteristics of the time series of the transformer oil temperature network change. In order to verify the effectiveness of the state visibility graph algorithm, three different types of time series are used in the experiment to construct the state visibility graph of the transformer oil temperature network. The three time series are periodic series, random series and fractal series.

[0136] The structural characteristics of different networks are analyzed by degree distribution. If the transformer oil temperature network state visibility graphs constructed with these three different types of time series have different network structures, it shows that the state visibility graph algorithm in this paper is effective.

[0137] The artificially generated periodic time series is used in the periodic time series verification. Let the period of the time series be 8, and the data sequence in each period be: {2,3,4,5,4,3,2,1}, so the corresponding state sequence is S = {(γ1,δ1), (γ2,δ1), (γ3,δ1), (γ4,δ4), (γ3,δ2), (γ2,δ2), (γ1,δ2), (γ0,δ3), (γ0,δ2)}, where there are 9 nodes in S, because when the period of the periodic sequence is greater than 1, the last element 1 of the periodic sequence has two local fluctuation characteristics, one is in the sequence, which is the minimum state, and the other is at the end of the sequence, which is a decreasing state according to the trend extension method. Figure 2 The state visibility graph is shown for time series with periods of 100, 200, and 500.

[0138] from Figure 2 From the changes in graphs a, b, and c, we can see that the structure of the state visible graph does not change with the increase in the number of cycles, indicating that the state visible graph of the periodic time series is a regular graph with a stable structure. When the time span of the periodic time series becomes longer, the number of nodes in the state visible graph remains unchanged, the weights between the edges continue to grow in a cycle, and the degree values ​​of all nodes remain in a relatively balanced state, so the degree distribution of the periodic time series is a Dirac distribution.

[0139] Two random sequences were used in the verification experiment of random time series. One is a random sequence X1 with a length of 10,000 and uniformly sampled from the interval [1,20); the other is a Gaussian random sequence with a mean of 10, a standard deviation of 1, and a length of 10,000.

[0140] Figure 3 The degree distribution of the random sequence X1 and the corresponding state X2 visible graph is plotted. Figure 3 As can be seen from Figure a and Figure b, the degree distribution curves of the two are similar, and the range of degree values ​​is relatively small. When the degree of the node gradually increases, the probability of degree distribution decreases rapidly, showing the structural characteristics of Poisson distribution, which is consistent with the degree distribution characteristics of random networks. At the same time, there is a form similar to exponential distribution at the tail of the distribution, indicating that there are a small number of central nodes with relatively large degrees in the network. The existence of these points causes the tail of the degree distribution to deviate from the Poisson distribution. This is because although the sample data of the transformer oil temperature change sampling is random as a whole, the random time series is an ordered set and has a certain correlation. In continuous sampling, it is a low-probability event to draw two large values ​​in a row, resulting in the appearance of a central node with relatively strong visibility, which in turn causes the degree distribution curve of the transformer oil temperature status visibility graph to deviate from the Poisson distribution curve at the tail.

[0141] In the fractal time series verification, the Conway sequence is selected as the experimental research object. The Conway sequence is a recursively generated fractal time series {a(t)-t / 2|t=1,2,...,n}, where a(t) is defined as follows:

[0142] a(t)=a(a(t-1))+a(ta(t-1)),n≥2

[0143] When t=1 or t=2, a(1)=a(2)=1.

[0144] The degree distribution of the Conway sequence consisting of 10,000 nodes and its corresponding state visible graph is as follows Figure 4 As shown. Figure 4 As can be seen in (b), the range of degree values ​​in the state visibility graph is relatively large, and as the degree value increases, the probability of degree distribution first decreases exponentially, and then when the degree value is large, the probability of degree distribution shows a trend of rising and then falling. This is because as the sequence continues to grow, new nodes with greater visibility continue to appear, and these nodes are relatively small. At the same time, the state visibility graph algorithm merges nodes with similar partial values ​​and similar local fluctuation states into one state node, resulting in a convex shape at the tail of the distribution curve.

[0145] To sum up, the degree distribution of the state visibility graphs of the three types of time series shows completely different characteristics, which shows that the state visibility graph effectively inherits some structural characteristics of the transformer oil temperature time series. By analyzing the structural characteristics of the state visibility graph, different types of time series can be accurately distinguished.

[0146] 4. Transformer oil temperature prediction algorithm based on state visibility graph

[0147] This experiment uses a prediction algorithm based on SVG to perform prediction analysis on the data sets ETTh1, PEMS-07, PEMS-BAY, METR-LA, and SOLAR. In order to verify the effectiveness of SVG in time series prediction, the present invention selected the ARIMA (Autoregressive Integrated Moving Average model) algorithm, DFNPM, and three different visible graph-based prediction algorithms in the comparative experiment, namely Zhang's visible graph prediction model, Li's visible graph prediction model, and Mao's visible graph prediction model. In the experiment, the sliding window method is used to select the training set, and the recursive prediction method is adopted for prediction. The experimental content is divided into two parts: a single-step forward prediction experiment and a multi-step forward prediction experiment.

[0148] (1) One-step forward prediction

[0149] In the single-step forward prediction experiment, the initial window length L is defined as 100, and the window moving step is defined as 1. Let the time series X = {(t1,y1), (t2,y2), ..., (t n ,y n )}, the specific experimental steps are as follows:

[0150] Step 1: Define an initial window of length L and move the window starting from the head of sequence X.

[0151] Step 2: Train the model using the data in the window.

[0152] Step 3: Calculate the predicted value using the model obtained in step 2

[0153] Step 4: Move the window to the right by a step length of step, and return to step 2. When the window moves to the end of the sequence, the prediction process ends.

[0154] Table 2 plots the experimental results of various prediction models on the datasets ETTh1, METR-LA, PEMS-07, PEMS-BAY, and SOLAR. The results with the smallest error are shown in bold, indicating the highest prediction accuracy.

[0155] On the data set ETTh1, the SVG model proposed by the present invention achieved better results than the other five models in the four evaluation indicators of MAE, MSE, RMSE and MAPE, proving that our invention can accurately predict the oil temperature of the transformer. In terms of the MSPE indicator, the prediction accuracy of SVG is lower than that of the visible graph prediction algorithm proposed by Zhang, Mao and Liu, indicating that the prediction error of a small number of nodes in the SVG model on the ETTh1 data set is higher than that of the three visible graph algorithms. However, after this part of the error is standardized as the percentage error MAPE, the SVG model is still in the optimal state, indicating that this small part of the prediction value with large error only accounts for a very small part of it.

[0156] On the datasets PEMS-07, PEMS-BAY and SOLAR, the results of the five evaluation indicators of the SVG model are all the lowest values, indicating that the performance of the SVG model on the three datasets PEMS-07, PEMS-BAY and SOLAR is better than that of the other five models, and the predicted values ​​on most of the data points are closer to the true values, proving that the model of the present invention can be widely promoted in other fields.

[0157] Table 2

[0158]

[0159] In summary, the state visibility graph algorithm performs stably in the recursive single-step forward prediction, has high prediction accuracy for transformer oil temperature changes, and as the amount of time series data increases, the prediction accuracy is better than other compared algorithms in different indicators.

[0160] (2) Multi-step Forward Prediction

[0161] Each prediction in multi-step forward prediction requires multiple prediction values. In the multi-step forward prediction experiment, this paper uses the recursive prediction method to predict the results when the step length is 2, 3, and 10, and studies the multi-step prediction performance of the model. The steps of multi-step forward prediction are as follows:

[0162] Step 1: Define an initial window of length L and a prediction step of step, and move the window from the beginning of the sequence.

[0163] Step 2: Train the model using the data in the window.

[0164] Step 3: Calculate the predicted value using the model obtained in step 2

[0165] Step 4: Add the predicted values ​​obtained in step 3 to the window and return to step 2. When the number of predicted values ​​added to the window is equal to the given prediction step, a multi-step forward prediction ends.

[0166] Step 5: Delete the value added to the window in step 4 from the window, move the window to the right by a distance of step, and return to step 2. When the window moves to the end of the sequence, the prediction process ends.

[0167] Table 3 shows the multi-step forward prediction results of each model on the dataset, where the best results are shown in bold.

[0168] Table 3

[0169]

[0170]

[0171] On the transformer oil temperature change data set ETTh1, the ARIMA algorithm performs best in 2-step, 3-step and 10-step forward predictions on the MSE and RMSE indicators and 10-step forward on the MAE indicator, but the SVG model performs second only to the best performance in the MAE, MSE and RMSE indicators in the 2-step, 3-step and 10-step forward predictions, and performs better than the ARIMA algorithm in the MAPE and MSPE indicators, indicating that the SVG model is better than the ARIMA algorithm in predicting the small value part of the time series. And from Figure 1In Figure a, we can see that the long-term trend of ETTh1 data is gradually decreasing over time, indicating that in the process of model training, as the amount of data increases, the prediction accuracy of the SVG model is also continuously improving. This is because in multi-step forward prediction, the state visible graph only needs to continuously add new nodes to the model, and the previous historical data information can be used in the later prediction process.

[0172] From Table 3, we can see that most of the evaluation indicators of the SVG model on the four datasets of PEMS-07, PEMS-BAY, METR-LA and SOLAR are better than those of other comparison models. In the datasets of PEMS-07 and PEMS-BAY, all indicators of the SVG model are better than those of other comparison models, which proves the robustness of the SVG model.

[0173] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A transformer oil temperature prediction method based on state visibility graph, characterized in that: The following steps are involved: Acquire an initial time-state sequence of transformer oil temperature based on the time series, wherein the initial time-state sequence includes several elements, including: time data and node state data; The initial time-state sequence of the transformer oil temperature is subjected to coarse-graining and local fluctuation characteristic analysis in sequence to obtain a state sequence of the time series; constructing a state visibility graph based on the state sequence of the time series; The node status at the current moment is obtained, and the oil temperature value of the transformer at the next moment is predicted based on the state visibility graph and the node status at the current moment.

2. The transformer oil temperature prediction method based on state visibility graph according to claim 1 is characterized in that: The process of processing and transforming the initial time-state sequence to construct a state visibility graph includes: Calculating the first-order mean absolute difference of the initial time-state sequence to obtain a first-order mean absolute difference value; defining a plurality of symbols based on the first-order mean absolute difference value; After the elements in the initial time-state sequence are coarse-grained, they are combined with the symbols to construct a symbol sequence based on the element mapping relationship; Converting the symbol sequence into an oil temperature state sequence; The state visible graph is obtained by converting and calculating the oil temperature state sequence based on a sliding visible graph algorithm.

3. The transformer oil temperature prediction method based on state visibility graph according to claim 2 is characterized in that: The calculation formula for calculating the first-order mean absolute difference value is: In the formula, R average represents the first-order mean absolute difference, y i Represents the transformer oil temperature value of the i-th node.

4. The transformer oil temperature prediction method based on state visibility graph according to claim 3 is characterized in that: The calculation formula for defining several symbols based on the first-order mean absolute difference value is: In the formula, k represents the total number of matches, y min Represents the minimum node state data in the initial time-state sequence, y max Represents the maximum node state data in the initial time-state sequence.

5. The transformer oil temperature prediction method based on state visibility graph according to claim 4 is characterized in that: The element mapping relationship is: Yes min +(m-1)R average ≤y t and y t <y min +mR average In the formula, γ m Represents the mth symbol.

6. The transformer oil temperature prediction method based on state visibility graph according to claim 5 is characterized in that: The local fluctuation characteristics include: a steady state, a rising state, a falling state, a maximum state and a minimum state.

7. The transformer oil temperature prediction method based on state visibility graph according to claim 6 is characterized in that: The process of predicting the oil temperature value of the transformer at the next moment based on the state visibility graph and the node state at the current moment includes: Calculate the state node of the node state at the current moment of observation in the oil temperature state sequence; Obtaining an in-degree node set and an out-degree node set of the state node in the state visible graph; Performing weighted summation on the out-degree node set to obtain an out-degree prediction value; The in-degree node set is calculated using a trend extension method to obtain an in-degree prediction value; The out-degree prediction value and the in-degree prediction value are weightedly summed to obtain the transformer oil temperature prediction value at the next moment.

8. The transformer oil temperature prediction method based on state visibility graph according to claim 7 is characterized in that: The calculation formula of the transformer oil temperature prediction value at the next moment is: In the formula, Indicates the predicted value of transformer oil temperature at the next moment, represents the predicted in-degree value, Indicates the predicted value of degree,

Images