Icing thickness prediction method and device, computer equipment, readable storage medium and program product

Abstract

The invention relates to an icing thickness prediction method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring a time sequence data set, wherein the time sequence data set comprises air temperature and humidity; inputting the air temperature and the humidity into an icing thickness target prediction model, and obtaining icing thickness prediction data output by the icing thickness target prediction model; the icing thickness target prediction model is generated by training a model comprising a convolutional neural network, a long-short-term memory network and an attention mechanism based on a training set. By adopting the method, the icing thickness prediction precision and accuracy can be improved.

Description

Technical Field

[0001] This application relates to the technical field of power grids, and in particular, to a method, device, computer device, computer-readable storage medium, and computer program product for predicting ice coating thickness. Background Art

[0002] Icing on power facilities has always been an important natural factor threatening the safe and stable operation of power grids. The occurrence range, frequency, and intensity of various extreme and severe weather conditions have all shown a significant increasing trend, and the harm caused to power grid operation has also shown a rapid upward trend. Monitoring the ice coating thickness of cables in real time and removing ice in a timely manner can reduce fault accidents such as line galloping, strand twisting, wire breakage, and tower collapse caused by icing, and reduce losses.

[0003] Currently, common methods for ice coating prediction are mostly based on physical model calculation methods or regression algorithms. Some introduce machine learning methods to model and predict the ice coating thickness. However, due to the single data sample and the problem of the ice coating reconstruction prediction model structure, there is a phenomenon of generally underestimating the reconstructed ice coating thickness.

[0004] However, the current ice coating thickness prediction methods still have the problem of insufficient accuracy in ice coating thickness prediction. Summary of the Invention

[0005] Based on this, it is necessary to provide an ice coating thickness prediction method, device, computer device, computer-readable storage medium, and computer program product that can improve the accuracy of ice coating thickness prediction for the above technical problems.

[0006] In a first aspect, this application provides an ice coating thickness prediction method, and the method includes:

[0007] Obtain a time series data set, where the time series data set includes air temperature and humidity;

[0008] Input the air temperature and the humidity into an ice coating thickness target prediction model, and obtain ice coating thickness prediction data output by the ice coating thickness target prediction model; the ice coating thickness target prediction model is generated by training a model including a convolutional neural network, a long short-term memory network, and an attention mechanism based on a training set.

[0009] In one embodiment, the training method of the ice coating thickness target prediction model includes:

[0010] Obtain a time series data sample set;

[0011] Process the time series data sample set based on a convolutional neural network unit, a long short-term memory network unit, and an attention mechanism unit to obtain a key feature vector;

[0012] Train the initial icing thickness prediction model using the key feature vector to obtain the target icing thickness prediction model.

[0013] In one embodiment, processing the time series data sample set by the convolutional neural network unit, the long short-term memory network unit and the attention mechanism unit to obtain the key feature vector includes:

[0014] Process the time series data sample set through the convolutional neural network unit to obtain the spatial features of the time series data sample set;

[0015] Extract the spatio-temporal features of the time series data sample set through the long short-term memory network unit;

[0016] Process the spatio-temporal features through the attention mechanism unit to obtain the key feature vector.

[0017] In one embodiment, the convolutional neural network unit includes a convolutional layer and a pooling layer; processing the time series data sample set through the convolutional neural network unit to obtain the spatial features of the time series data sample set includes:

[0018] Perform convolutional processing on the time series data sample set through the convolutional layer to obtain a data sequence;

[0019] Perform pooling processing on the data sequence through the pooling layer to obtain the spatial features.

[0020] In one embodiment, processing the spatio-temporal features through the attention mechanism unit to obtain the key feature vector includes:

[0021] Determine the attention score according to the obtained first weight matrix and the first bias matrix; the first weight matrix is the weight matrix from the input layer to the hidden layer of the attention mechanism unit; the first bias matrix is the bias matrix of the attention mechanism unit;

[0022] Determine the weight corresponding to the spatio-temporal features according to the attention score;

[0023] Determine the key feature vector according to the spatio-temporal features and the corresponding weights.

[0024] In one embodiment, after obtaining the predicted icing thickness data output by the target icing thickness prediction model, the method includes:

[0025] Obtain the actual icing thickness data corresponding to the temperature and humidity in the time series dataset;

[0026] Determine an evaluation result based on the ice thickness prediction data and the actual ice thickness data, where the evaluation result includes the mean squared error, mean absolute error, and mean error magnitude between the ice thickness prediction data and the actual ice thickness data;

[0027] When the mean squared error, the mean absolute error, and the mean error magnitude are all greater than their corresponding thresholds, it indicates that the ice thickness target prediction model meets the prediction accuracy requirements;

[0028] When any one of the mean squared error, the mean absolute error, and the mean error magnitude is less than or equal to its corresponding threshold, adjust the hyperparameters of the ice thickness target prediction model to obtain a new ice thickness target prediction model.

[0029] In a second aspect, the present application further provides an ice thickness prediction device, which includes:

[0030] An acquisition module for acquiring a time series data set, where the time series data set includes temperature and humidity;

[0031] A processing module for inputting the temperature and the humidity into the ice thickness target prediction model and obtaining the ice thickness prediction data output by the ice thickness target prediction model; the ice thickness target prediction model is generated by training a model including a convolutional neural network, a long short-term memory network, and an attention mechanism based on a training set.

[0032] In a third aspect, the present application further provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0033] In a fourth aspect, the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0034] In a fifth aspect, the present application further provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0035] The above-mentioned icing thickness prediction method, device, computer equipment, computer-readable storage medium and computer program product first obtain a time-series data set, which includes air temperature and humidity; secondly, input the air temperature and humidity into the target icing thickness prediction model, and obtain the icing thickness prediction data output by the target icing thickness prediction model; since the target icing thickness prediction model is trained based on a training set for a model including a convolutional neural network, a long short-term memory network and an attention mechanism, which combines multiple models such as the convolutional neural network and the long short-term memory network, and the attention mechanism continuously corrects the initial prediction result, it can improve the prediction accuracy. Description of the Drawings

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments of the present application or related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0037] Figure 1 It is an application environment diagram of the icing thickness prediction method in an embodiment;

[0038] Figure 2 It is a flowchart of the icing thickness prediction method in an embodiment;

[0039] Figure 3 It is a flowchart of the training method of the target icing thickness prediction model in an embodiment;

[0040] Figure 4 It is a flowchart of the generation method of the key feature vector in an embodiment;

[0041] Figure 5 It is a flowchart of the generation method of the spatial features of the time-series data sample set in an embodiment;

[0042] Figure 6 It is a flowchart of the generation method of the key feature vector in another embodiment;

[0043] Figure 7 It is a flowchart of the evaluation method of the target icing thickness prediction model in an embodiment;

[0044] Figure 8 It is a structural block diagram of the icing thickness prediction device in an embodiment;

[0045] Figure 9 It is an internal structure diagram of the computer equipment in an embodiment. Detailed Embodiments

[0046] In order to make the objectives, technical solutions, and advantages of the present application more clearly understood, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0047] The icing thickness prediction method provided by the embodiments of the present application can be applied to an application environment as Figure 1 shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or can be placed in the cloud or other network servers. The server obtains a time series data set, and the time series data set includes air temperature and humidity; inputs the air temperature and humidity into the icing thickness target prediction model, and obtains the icing thickness prediction data output by the icing thickness target prediction model; the icing thickness target prediction model is generated by training a model including a convolutional neural network, a long short-term memory network, and an attention mechanism based on a training set. Among them, the terminal 102 can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The head-mounted device can be a virtual reality (VR) device, an augmented reality (AR) device, smart glasses, etc. The server 104 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0048] In an exemplary embodiment, as Figure 2 shown, a method for predicting icing thickness is provided. Taking the method applied to the Figure 1 server in it as an example, it includes the following steps S202 to step S204.

[0049] Among them:

[0050] Step S202, obtain a time series data set, and the time series data set includes air temperature and humidity.

[0051] Among them, the time series data set includes the air temperature and humidity at the time step, which can be denoted as X t , representing the data set at the t-th time step. X t = [X t,气温 , X t,湿度 . Among them, X t,气温 represents the air temperature at the t-th time step, and X t,湿度 represents the humidity at the t-th time step.

[0052] Optionally, the server obtains the dataset X of the time step t = t,气温 , t,湿度 .

[0053] Step S204: Input the air temperature and humidity into the icing thickness target prediction model, and obtain the icing thickness prediction data output by the icing thickness target prediction model.

[0054] Among them, the icing thickness target prediction model is generated by training a model including a convolutional neural network, a long short-term memory network, and an attention mechanism based on a training set.

[0055] Optionally, the server pre-trains the icing thickness target prediction model. In actual operation, the server trains a model including a convolutional neural network, a long short-term memory network, and an attention mechanism according to the time series data sample set as the training set to obtain the icing thickness target prediction model. Among them, the convolutional neural network is used to extract features from the time series data sample set to capture spatial dependence and obtain spatial features; then the long short-term memory network is used to process the spatial features to capture time dependence and obtain spatio-temporal features, and the icing thickness prediction data is output according to the spatio-temporal features.

[0056] Optionally, the server obtains the dataset of a certain time step, including the air temperature and humidity, and inputs the dataset of the certain time step into the pre-trained icing thickness target prediction model. The icing thickness target prediction model processes the dataset of the certain time step and outputs the icing thickness prediction data corresponding to the certain time step. For example, the server obtains the dataset X t = t,气温 , t,湿度 of the t-th time step, and inputs X t = t,气温 , t,湿度 into the pre-trained icing thickness target prediction model. The icing thickness target prediction model processes the dataset of the t-th time step and outputs the icing thickness prediction data corresponding to the t-th time step.

[0057] In the above icing thickness prediction method, first, a time series dataset is obtained, and the time series dataset includes the air temperature and humidity; secondly, the air temperature and humidity are input into the icing thickness target prediction model, and the icing thickness prediction data output by the icing thickness target prediction model is obtained; since the icing thickness target prediction model is generated by training a model including a convolutional neural network, a long short-term memory network, and an attention mechanism based on a training set, the convolutional neural network and the long short-term memory network are combined in multiple models, and the attention mechanism continuously corrects the initial prediction result to improve the prediction accuracy.

[0058] In an exemplary embodiment, as Figure 3 shown, the training method of the ice thickness target prediction model includes steps S302 to S306. Among them:

[0059] Step S302: Obtain a time series data sample set for ice thickness prediction.

[0060] Each sample data in the time series data sample set includes sample humidity, sample temperature, and the actual value of the ice thickness. Optionally, the server obtains data such as humidity, temperature, and ice thickness collected by each endpoint, supplements blank values using the average interpolation method, performs normalization processing to keep the data dimensions consistent, constructs a time series training set of the first duration and extrapolates a time series training set of the second duration. For example, extrapolate a 12-hour time series training set from a 124-hour time series training set. Randomly divide the time series data set into a training set and a test set in a ratio of 8:2. Among them, the time series training set of the first duration is the data obtained in reality, and the time series training set of the second duration is extrapolated from the time series training set of the first duration.

[0061] The server performs anomaly identification on the temperature, humidity, and ice thickness monitored by each endpoint every hour. The correct threshold is shown in Table 1. Values outside the correct threshold are identified as outliers and removed after identification. After removing the abnormal data, the energy consumption value at this moment becomes a missing value. Therefore, the missing value needs to be filled in using the adjacent interpolation method to make it as close to the true value as possible. Finally, the data is normalized to eliminate the adverse effects caused by singular sample data.

[0062] Table 1 Input categories and correct thresholds

[0063]

[0064]

[0065] The data after filling is further normalized according to formula (1) to obtain a time series data sample set, denoted as X = {x1, x2, …, x t}.

[0066]

[0067] Among them, x is the monitored data before normalization; x′ is the monitored data after normalization; x min is the minimum value of the monitored data set; x max is the maximum value of the monitored data. This formula applies to all monitored data (temperature, humidity, ice thickness) because each parameter has its own minimum and maximum values, and the normalized data will be unified to the same scale for subsequent processing.

[0068] Step S304: Process the time-series data sample set based on the convolutional neural network unit, long short-term memory network unit, and attention mechanism unit to obtain the key feature vector.

[0069] The server uses an attention-based CNN-LSTM hybrid model as the initial ice thickness prediction model. The server processes the time-series data sample set based on the convolutional neural network unit, long short-term memory network unit, and attention mechanism unit to obtain the key feature vector c.

[0070] Step S306: Use the key feature vector to train the initial ice thickness prediction model to obtain the target ice thickness prediction model.

[0071] Optionally, the server uses the key feature vector c to further train the initial ice thickness prediction model to obtain the target ice thickness prediction model. When inputting the prediction data set, the target ice thickness prediction model outputs the ice thickness prediction data.

[0072] In this embodiment, the initial ice thickness prediction model is further revised and adjusted through the key feature vector to obtain the target ice thickness prediction model, thereby improving the accuracy of the ice thickness prediction result.

[0073] In an exemplary embodiment, as Figure 4 shown, processing the time-series data sample set based on the convolutional neural network unit, long short-term memory network unit, and attention mechanism unit to obtain the key feature vector includes steps S402 to S406. Wherein:

[0074] Step S402: Process the time-series data sample set through the convolutional neural network unit to obtain the spatial features of the time-series data sample set.

[0075] Optionally, the server processes the time-series data sample set through the convolutional neural network unit to capture the spatial dependence and obtain the spatial features of the time-series data sample set.

[0076] Step S404: Extract the spatio-temporal features of the time-series data sample set through the long short-term memory network unit.

[0077] Optionally, the server extracts the spatial features through the long short-term memory network unit to capture the time dependence and ensure the coherence of the data stream, and obtains the spatio-temporal features of the time-series data sample set.

[0078] Step S406: Process the spatio-temporal features through the attention mechanism unit to obtain the key feature vector.

[0079] Optionally, the spatio-temporal feature vector V output by the server through the long short-term memory network unit LSTM enters the attention mechanism layer. Through the attention mechanism, the weight β of the feature vector at each time step is calculated i , and the key feature vector c is extracted.

[0080] In this embodiment, the initial icing thickness prediction model can be further trimmed by the extracted key feature vector to improve the accuracy of the icing thickness data prediction.

[0081] In an exemplary embodiment, as Figure 5 shown, the convolutional neural network unit includes a convolutional layer and a pooling layer; the spatio-temporal data sample set is processed by the convolutional neural network unit to obtain the spatial features of the spatio-temporal data sample set, including steps S502 to S504. Among them:

[0082] Step S502, perform convolutional processing on the spatio-temporal data sample set through the convolutional layer to obtain a data sequence.

[0083] The convolutional neural network unit belongs to a one-dimensional convolutional neural network, and the input parameters are the spatio-temporal data sample set after normalization processing, including air temperature, humidity, and icing thickness data, denoted as X = {x1, x2,..., x t}. Where x t = [x t,气温 , x t,湿度 , x t,覆冰厚度 represents the input data at the t-th time step.

[0084] Among them, the mathematical expression of the convolutional layer is shown in formula (2).

[0085]

[0086] In the formula, k represents the number of convolutional operations; N is the length of u.

[0087] y(k) represents the output sequence of the convolutional operation; the physical meaning represents the eigenvalue obtained after the k-th convolutional operation.

[0088] h(k) represents the convolutional kernel (filter) sequence, and the physical meaning represents the value of the convolutional kernel at the k-th position.

[0089] u(k) represents the input data sequence, and the physical meaning represents the value of the input data at the k-th position.

[0090] i represents the summation index variable, and the physical meaning represents the relative position of the current calculation in the convolutional operation.

[0091] u(i) represents the value of the input data sequence at the i-th position, and the physical meaning represents the value of the input data at the relative position i.

[0092] h(k - i) represents the value of the convolutional kernel sequence at the relative position k - i; its physical meaning indicates the value of the convolutional kernel at the relative position k - i.

[0093] Among them, the type of y(k) is a scalar (a single value) or a vector (depending on the specific application). Its role is the feature extracted by the convolutional layer and is used for subsequent network processing.

[0094] The type of h(k) is a vector (usually a sequence of fixed length). Its role is that the convolutional kernel is used to extract features from the input data, and its value is learned during the training process.

[0095] The type of u(k) is a vector (a sequence matching the length of the convolutional kernel). Its role is the input of the convolutional operation, usually data such as temperature, humidity, ice thickness, etc. that has been pre - processed (such as normalized).

[0096] The type of i is an integer (ranging from 1 to N). Its role is to traverse the corresponding positions of the convolutional kernel and the input data to calculate the convolutional result.

[0097] The type of u(i) is a scalar (a single value). Its role is to multiply with the value h(k - i) of the convolutional kernel and participate in the convolutional calculation.

[0098] The type of h(k - i) is a scalar (a single value). Its role is to multiply with the value u(i) of the input data and participate in the convolutional calculation.

[0099] Step S504, perform pooling processing on the data sequence through the pooling layer to obtain spatial features.

[0100] The mathematical representation of the pooling layer is shown in formula (3).

[0101]

[0102] In the formula, is the output value of the pooling layer, and its physical meaning indicates the pooling result at the position (i, j) on the k - th feature map of the l - th layer.

[0103] x and y represent the local indices within the pooling window, and their physical meaning indicates the relative positions of each element within the pooling window.

[0104] i and j represent the position indices of the pooling output. Their physical meaning indicates the position (i, j) of the pooling output feature map.

[0105] s0 represents the stride, and its physical meaning indicates the stride at which the pooling window slides on the input feature map.

[0106] f represents the size of the pooling window, and its physical meaning indicates the width and height of the pooling window.

[0107] Let \(p\) denote the pooling type parameter. When \(p\rightarrow+\infty\), the maximum value within the pooling region is taken, i.e., max pooling.

[0108] Among them, is of scalar (single value) type. Its function is that the pooling operation aggregates the local regions of the input feature map (such as taking the maximum value or the average value), reduces the data dimension, and retains important features.

[0109] Both \(x\) and \(y\) are of integer type (the value range is from 1 to \(f\)). Their function is to traverse all elements within the pooling window and calculate the aggregated value.

[0110] Both \(i\) and \(j\) are of integer type. Their function is to locate the output position of the pooling operation.

[0111] \(s_0\) is of integer type. Its function is to control the size of the pooling output. The larger the stride, the smaller the size of the output feature map.

[0112] \(f\) is an integer. Its function is to define the size of the local region of the pooling operation.

[0113] \(p\) is of real number type. Its function is to control the type of the pooling operation.

[0114] The pooling operation slides a window of a fixed size (\(f\times f\)) on the input feature map, aggregates the values within the window (such as taking the maximum value or the average value), and takes the result as the value of the output feature map. The specific steps are as follows: For each position \((i, j)\) of the output feature map, locate the corresponding region on the input feature map: The upper left corner of the region is \((s oi , s oj ); Traverse all positions \((x, y)\) within the pooling window and calculate the aggregated value. Take the aggregated value as the value of the output feature map at position \((i, j)\).

[0115] Optionally, the convolutional neural network unit further includes a connection layer; after the pooling layer performs pooling processing on the data sequence to obtain spatial features, the spatial features are passed to the long short-term memory network unit through the connection layer; the long short-term memory network unit includes an input gate, a forget gate, and an output gate; the spatial features include the spatial features of the current time step; the long short-term memory network unit extracts the spatio-temporal features of the time series data sample set, including: obtaining the hidden state of the previous time step; processing the spatial features of the current time step and the hidden state of the previous time step through the input gate to obtain the information to be added; updating to obtain the current cell state through the input gate and the forget gate based on the information to be added and the candidate cell state, where the candidate cell state is determined based on the second weight matrix and the second bias matrix, the second weight matrix is the weight matrix of the candidate cell state; the second bias matrix is the bias matrix of the candidate cell state; outputting the spatio-temporal features according to the current cell state through the output gate.

[0116] Further, the features extracted by the server through the convolutional neural network unit CNN are transmitted to the long short-term memory network (LSTM) through the connection layer Connect. The LSTM layer extracts the spatio-temporal features of the icing data through the calculations of the input gate, forget gate, and output gate, i.e., formulas (4)-(9).

[0117] i t = σ(W i ·[h t-1 , x t + b i ) Formula (1)

[0118] where i t is the output of the input gate, with dimension (H); W i and b i are the weight matrix and bias of the input gate, with dimensions (H, H+D) and (H) respectively; h t-1 is the hidden state of the previous time step, with dimension (H); x t is the input of the current time step. σ represents the sigmoid function, which is used to generate a value between 0 and 1, indicating which values in each cell state should be updated.

[0119] Then, through the multiplication operation, the input gate determines which parts of the new input information should be added to the cell state and passes the result to the cell state:

[0120]

[0121] where is the candidate cell state, with dimension (H); W C has dimension (H, H+D), and b C has dimension (H), which are the weight matrix and bias of the candidate cell state respectively.

[0122] The cell state is updated through the input gate and the forget gate:

[0123]

[0124] where C t is the current cell state; f t is the output of the forget gate; C t-1 represents the cell state of the previous time step.

[0125] C t-1 carries the long-term memory information from time step 1 to t-1 and is updated to C t at time step t.

[0126] Next, the data is passed into the forget gate, which generates a value between 0 and 1 through the sigmoid function, indicating which information will be forgotten or retained:

[0127] f t =σ(W f ·[h t-1 ,x t +b f ) Equation (4)

[0128] In the formula, f t is the output of the forget gate, with a dimension of (H); W f and b f are the weight matrix and bias of the forget gate, with dimensions of (H, H+D) and (H), respectively.

[0129] The output gate generates the hidden state at the current time step through the sigmoid function and the tanh function:

[0130] o t =σ(W o ·[h t-1 ,x t +b o ) Equation (5)

[0131] h t =o t tanh(C t ) Equation (6)

[0132] In the formula, o t is the output of the output gate, with a dimension of (H); h t is the hidden state at the current time step, with a dimension of (H); W o and b o are the weight matrix and bias of the output gate, with dimensions of (H, H+D) and (H), respectively.

[0133] The hyperparameters selected in this application and their values are shown in the following table.

[0134] Table 2 Hyperparameters of the LSTM model

[0135]

[0136] In this embodiment, the input data is subjected to feature extraction through the CNN unit to capture spatial dependence. The connection layer (Connect layer) transfers the features extracted by the CNN unit to the LSTM unit to ensure the coherence of the data stream. The LSTM unit processes the time-series data to capture time dependence, which can improve the prediction accuracy and precision of the icing data.

[0137] In an exemplary embodiment, such asFigure 6 As shown, the spatio-temporal features are processed by the attention mechanism unit to obtain the key feature vector, including steps S602 to S606. Among them:

[0138] Step S602: Determine the attention score according to the obtained first weight matrix and first bias matrix.

[0139] Among them, the first weight matrix is the weight matrix from the input layer to the hidden layer of the attention mechanism unit; the first bias matrix is the bias matrix of the attention mechanism unit.

[0140] Optionally, the first weight matrix and first bias matrix obtained by the server are used to determine the attention score through the attention score function and formula (10).

[0141] e i = a(v i ) = σ(W i ·v i + b) Formula (10)

[0142] Among them, a(·) is the attention score function at the i-th time step. W i is the weight matrix from the input layer to the hidden layer, b is the bias value matrix, and e i is the attention score.

[0143] Step S604: Determine the weight corresponding to the spatio-temporal features according to the attention score.

[0144] Optionally, the server determines the weight β corresponding to the spatio-temporal features according to the attention score through an activation function, such as calculating according to the softmax function and formula (11). i .

[0145]

[0146] Among them, exp(·) is the exponential function; e i and e k are the attention scores.

[0147] Step S606: Determine the key feature vector according to the spatio-temporal features and the corresponding weights.

[0148] Optionally, let the feature vector output by the LSTM unit be V = {v1, v2,..., v m}}. Among them, m is the total number of time steps, and v i is the feature vector at the i-th time step. The feature vector V output by the LSTM layer enters the attention mechanism layer. The attention mechanism calculates the weight β of each time step feature vector i, extract the key feature vector c. The calculation of the attention score is shown in formula (12), and the feature vector is weighted through the sigmoid function. The server determines the key feature vector c based on the spatio-temporal feature V and the corresponding weight β i , and determines the key feature vector c according to formula (12).

[0149]

[0150] In this embodiment, the key feature vector is determined through the attention mechanism and the spatio-temporal vector output by LSTM, so as to trim the initial icing thickness prediction model to improve the prediction accuracy of the icing thickness data.

[0151] In an exemplary embodiment, as Figure 7 shown, after obtaining the icing thickness prediction data output by the icing thickness target prediction model, the evaluation method of the icing thickness target prediction model includes steps S702 to S706. Among them:

[0152] Step S702, obtain the actual icing thickness data corresponding to the temperature and humidity in the time series dataset.

[0153] Optionally, the server obtains the actual icing thickness data corresponding to the temperature and humidity in the time series dataset.

[0154] Step S704, determine the evaluation result based on the icing thickness prediction data and the actual icing thickness data. The evaluation result includes the mean squared error, mean absolute error, and mean error magnitude between the icing thickness prediction data and the actual icing thickness data.

[0155] Among them, the mean squared error is denoted as MSE; the mean absolute error is denoted as MAE; the mean error magnitude is denoted as RMSE. MSE is used to measure the mean squared error between the predicted value and the actual value, and can effectively reflect the prediction accuracy of the model. MAE is used to measure the mean absolute error between the predicted value and the actual value, and can more robustly reflect the prediction error of the model and is not sensitive to outliers. RMSE is the square root of MSE and is used to measure the mean error magnitude between the predicted value and the actual value. Its unit is the same as the actual value, which is convenient for intuitively understanding the prediction performance of the model.

[0156] The data test and evaluation adopt three indicators of MSE, MAE, and RMSE. Among them, the calculation formula of the mean squared error MSE is shown in formula (13).

[0157]

[0158] Among them, y i represents the actual icing thickness value of the i-th sample, represents the icing thickness value predicted by the model, and n is the number of samples in the test set.

[0159] The calculation formula for the mean absolute error denoted as MAE is shown in Formula (14).

[0160]

[0161] Among them, y i represents the actual ice accretion thickness value of the i-th sample, represents the ice accretion thickness value predicted by the model, and n is the number of samples in the test set.

[0162] The calculation formula for the root mean square error RMSE is shown in Formula (15).

[0163]

[0164] Among them, y i represents the actual ice accretion thickness value of the i-th sample, represents the ice accretion thickness value predicted by the model, and n is the number of samples in the test set.

[0165] Step S706, when the mean squared error, mean absolute error, and root mean square error are all greater than their corresponding thresholds, it indicates that the ice accretion thickness target prediction model meets the prediction accuracy requirements; when any one of the mean squared error, mean absolute error, and root mean square error is less than or equal to the corresponding threshold, the hyperparameters of the ice accretion thickness target prediction model are adjusted to obtain a new ice accretion thickness target prediction model.

[0166] According to the calculation results of MSE, MAE, and RMSE, analyze the prediction performance of the model on the test set. Lower MSE, MAE, and RMSE values indicate that the model has higher prediction accuracy and stability. According to the evaluation results, further optimize the model structure or adjust the hyperparameters until the performance of the model on the validation set and the test set reaches the expected goal.

[0167] Optionally, when the mean squared error, mean absolute error, and root mean square error are all greater than their corresponding thresholds, it indicates that the ice accretion thickness target prediction model meets the prediction accuracy requirements and no further correction is required.

[0168] When any one of the mean squared error, mean absolute error, and root mean square error is less than or equal to the corresponding threshold, the hyperparameters of the ice accretion thickness target prediction model are adjusted to obtain a new ice accretion thickness target prediction model.

[0169] In this embodiment, according to the calculation results of MSE, MAE, and RMSE, analyze the prediction performance of the model on the test set. According to the evaluation results, further optimize the model structure or adjust the hyperparameters until the performance of the model on the validation set and the test set reaches the expected goal.

[0170] In an exemplary embodiment, it is divided into two stages. The first stage is the training stage of the icing thickness target prediction model. The second stage is the usage stage of the icing thickness target prediction model.

[0171] The first stage: The server obtains data such as humidity, temperature, and icing thickness collected by each endpoint, supplements blank values using the mean interpolation method, performs normalization processing to keep the data dimensions consistent, constructs a time series training set of the first duration, and extrapolates a time series training set of the second duration. For example, extrapolate a 12-hour time series training set from a 124-hour time series training set. Randomly divide the time series data set into a training set and a test set in a ratio of 8:2. Among them, the time series training set of the first duration is the data obtained in real time, and the time series training set of the second duration is obtained by extrapolating from the time series training set of the first duration.

[0172] The server performs anomaly recognition on the temperature, humidity, and icing thickness monitored by each endpoint every hour. The correct threshold is shown in Table 1. Values outside the correct threshold are identified as outliers and removed after identification. After removing the abnormal data, the energy consumption value at this moment becomes a missing value. Therefore, the missing value needs to be filled in using the adjacent interpolation method to make it as close to the real value as possible. Finally, the data is normalized to eliminate the adverse effects caused by singular sample data.

[0173] Table 1 Input categories and correct thresholds

[0174]

[0175] The data after completion is further normalized according to formula (1) to obtain a time series data sample set, denoted as X = {x1, x2, …, x t}.

[0176]

[0177] Among them, x is the monitored data before normalization; x′ is the monitored data after normalization; x min is the minimum value of the monitoring data set; x max is the maximum value of the monitoring data. This formula is applicable to all monitored data (temperature, humidity, icing thickness) because each parameter has its own minimum and maximum values, and the normalized data will be unified to the same scale, which is convenient for subsequent processing.

[0178] The convolutional neural network unit includes a convolutional layer, a pooling layer, and a connection layer. The server belongs to a one-dimensional convolutional neural network through the convolutional neural network unit. The input parameters are the time series data sample set after normalization processing, including temperature, humidity, and icing thickness data, denoted as X = {x1, x2, …, x t}. Among them, x t = [x t,气温 , xt,湿度 , x t,覆冰厚度 represents the input data at the t-th time step.

[0179] Among them, the mathematical expression of the convolutional layer is shown in formula (2).

[0180]

[0181] The data sequence is pooled through the pooling layer to obtain spatial features. The mathematical representation of the pooling layer is shown in formula (3).

[0182]

[0183] The spatial features are passed to the long short-term memory network unit through the connection layer; the LSTM layer extracts the spatio-temporal features of the icing data through the calculations of the input gate, forget gate, and output gate, namely formulas (4)-(9).

[0184] i t = σ(W i · [h t-1 , x t + b i ) Formula (7)

[0185] Among them, i t is the output of the input gate, with dimension (H); W i and b i are the weight matrix and bias of the input gate, with dimensions (H, H + D) and (H) respectively; h t-1 is the hidden state of the previous time step, with dimension (H); x t is the input of the current time step. σ represents the sigmoid function, which is used to generate a value between 0 and 1, indicating which values in each cell state should be updated.

[0186] Then, through the multiplication operation, the input gate determines which parts of the new input information should be added to the cell state and passes the result to the cell state:

[0187]

[0188] Among them, is the candidate cell state, with dimension (H); W C has dimension (H, H + D), and b C has dimension (H), which are the weight matrix and bias of the candidate cell state respectively.

[0189] The cell state is updated through the input gate and forget gate:

[0190]

[0191] Among them, C t is the current cell state; f t is the output of the forget gate; C t-1 represents the cell state of the previous time step.

[0192] C t-1 carries the long-term memory information from time step 1 to t - 1 and is updated to C t at time step t.

[0193] Next, the data is passed into the forget gate, and the forget gate generates a value between 0 and 1 through the sigmoid function, indicating which information will be forgotten or retained:

[0194] f t = σ(W f · [h t-1 , x t + b f ) Equation (10)

[0195] In the formula, f t is the output of the forget gate, with a dimension of (H,); W f and b f are the weight matrix and bias of the forget gate respectively, with dimensions of (H, H + D) and (H,) respectively.

[0196] The output gate generates the hidden state of the current time step through the sigmoid function and the tanh function:

[0197] o t = σ(W o · [h t-1 , x t + b o ) Equation (11)

[0198] h t = o t tanh(C t ) Equation (12)

[0199] In the formula, o t is the output of the output gate, with a dimension of (H); h t is the hidden state of the current time step, with a dimension of (H); W o and b o are the weight matrix and bias of the output gate respectively, with dimensions of (H, H + D) and (H) respectively.

[0200] The hyperparameters selected in this application and their values are shown in the following table.

[0201] Table 2 Hyperparameters of the LSTM model

[0202]

[0203] The first weight matrix and the first bias matrix obtained by the server are used to determine the attention scores through the attention score function and formula (10).

[0204] e i = a(v i ) = σ(W i ·v i + b) Formula (10)

[0205] where a(·) is the attention score function at the i-th time step. W i is the weight matrix from the input layer to the hidden layer, b is the bias value matrix, and e i is the attention score.

[0206] Step S604: Determine the weights corresponding to the spatio-temporal features according to the attention scores.

[0207] Optionally, the server determines the weights β corresponding to the spatio-temporal features according to the attention scores by calculating through an activation function, such as the softmax function according to formula (11). i .

[0208]

[0209] where exp(·) is the exponential function; e i and e k are the attention scores.

[0210] Let the feature vector output by the LSTM cell be V = {v1, v2, …, v m}. Among them, m is the total number of time steps, and v i is the feature vector at the i-th time step. The feature vector V output by the LSTM layer enters the attention mechanism layer. The attention mechanism extracts the key feature vector c by calculating the weights β i of the feature vector at each time step. The calculation of the attention score is shown in formula (12), and the feature vector is weighted through the sigmoid function. The server determines the key feature vector c according to the spatio-temporal feature V and the corresponding weight β i according to formula (12).

[0211]

[0212] The server uses the key feature vector c to further train the initial icing thickness prediction model to obtain the target icing thickness prediction model. When inputting the prediction data set, the icing thickness prediction data output by the target icing thickness prediction model is obtained.

[0213] The server obtains the actual icing thickness data corresponding to the temperature and humidity in the time series dataset.

[0214] The mean squared error is denoted as MSE; the mean absolute error is denoted as MAE; the root mean squared error is denoted as RMSE. MSE is used to measure the mean squared error between the predicted value and the actual value, and can effectively reflect the prediction accuracy of the model. MAE is used to measure the mean absolute error between the predicted value and the actual value, can more robustly reflect the prediction error of the model, and is not sensitive to outliers. RMSE is the square root of MSE, and is used to measure the mean error amplitude between the predicted value and the actual value. Its unit is the same as the actual value, which is convenient for intuitively understanding the prediction performance of the model.

[0215] Obtain the actual ice coating thickness data corresponding to the air temperature and humidity in the time series dataset; the data test and evaluation adopt three indicators of MSE, MAE, and RMSE. Among them, the calculation formula of the mean squared error MSE is shown in formula (13).

[0216]

[0217] Among them, y i represents the actual ice coating thickness value of the i-th sample, represents the ice coating thickness value predicted by the model, and n is the number of samples in the test set.

[0218] The calculation formula of the mean absolute error denoted as MAE is shown in formula (14).

[0219]

[0220] Among them, y i represents the actual ice coating thickness value of the i-th sample, represents the ice coating thickness value predicted by the model, and n is the number of samples in the test set.

[0221] The calculation formula of the root mean squared error RMSE is shown in formula (15).

[0222]

[0223] Among them, y i represents the actual ice coating thickness value of the i-th sample, represents the ice coating thickness value predicted by the model, and n is the number of samples in the test set.

[0224] When the mean squared error, the mean absolute error, and the root mean squared error are all greater than the corresponding thresholds, it indicates that the ice coating thickness target prediction model meets the prediction accuracy requirements and does not need to be corrected again.

[0225] When any one of the mean squared error, the mean absolute error, and the root mean squared error is less than or equal to the corresponding threshold, the hyperparameters of the ice coating thickness target prediction model are adjusted to obtain a new ice coating thickness target prediction model.

[0226] The second stage: The server obtains the dataset X at the time step t =[X t,气温 , X t,湿度 . X t represents the dataset at the t-th time step. X t =[X t,气温 , X t,湿度 . Among them, X t,气温 represents the air temperature at the t-th time step, and X t,湿度 represents the humidity at the t-th time step. The server obtains the dataset X at the t-th time step t =[X t,气温 , X t,湿度 , and inputs X t =[X t,气温 , X t,湿度 into the pre-trained icing thickness target prediction model. The icing thickness target prediction model processes the dataset at the t-th time step and outputs the icing thickness prediction data corresponding to the t-th time step.

[0227] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the arrows, these steps do not necessarily have to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order restriction, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages do not necessarily have to be executed at the same moment, but can be executed at different moments. The execution order of these steps or stages does not necessarily have to be sequential, but can be executed alternately or alternately with at least some of the steps or stages in other steps or other steps.

[0228] Based on the same inventive concept, the embodiments of the present application also provide an icing thickness prediction device for implementing the above-mentioned icing thickness prediction method. The implementation solutions provided by this device to solve problems are similar to the implementation solutions described in the above method. Therefore, the specific limitations in one or more of the following embodiments of the icing thickness prediction device can refer to the limitations on the icing thickness prediction method in the above text, and will not be repeated here.

[0229] In an exemplary embodiment, as Figure 8 shown, an icing thickness prediction device is provided, including: an acquisition module 801 and a processing module 802, where:

[0230] The acquisition module 801 is used to acquire a time series dataset, and the time series dataset includes air temperature and humidity;

[0231] A processing module 802 is configured to input air temperature and humidity into an icing thickness target prediction model, and obtain icing thickness prediction data output by the icing thickness target prediction model. The icing thickness target prediction model is generated by training a model including a convolutional neural network, a long short-term memory network, and an attention mechanism based on a training set.

[0232] In an exemplary embodiment, an icing thickness prediction device is provided, further including a training module, configured to obtain a time series data sample set; process the time series data sample set based on a convolutional neural network unit, a long short-term memory network unit, and an attention mechanism unit to obtain a key feature vector; and train an initial icing thickness prediction model using the key feature vector to obtain an icing thickness target prediction model.

[0233] In an exemplary embodiment, the training module is further configured to process the time series data sample set through the convolutional neural network unit to obtain spatial features of the time series data sample set; extract the spatial features through the long short-term memory network unit to obtain spatio-temporal features of the time series data sample set; and process the spatio-temporal features through the attention mechanism unit to obtain a key feature vector.

[0234] In an exemplary embodiment, the convolutional neural network unit includes a convolutional layer and a pooling layer. The training module is further configured to perform convolutional processing on the time series data sample set through the convolutional layer to obtain a data sequence; and perform pooling processing on the data sequence through the pooling layer to obtain spatial features.

[0235] In an exemplary embodiment, the training module is further configured to determine an attention score according to the obtained first weight matrix and first bias matrix. The first weight matrix is the weight matrix from the input layer to the hidden layer of the attention mechanism unit. The first bias matrix is the bias matrix of the attention mechanism unit. Determine the weight corresponding to the spatio-temporal features according to the attention score; and determine the key feature vector according to the spatio-temporal features and the corresponding weights.

[0236] In an exemplary embodiment, an icing thickness prediction device is provided, further including an evaluation and adjustment module, configured to obtain the actual icing thickness data corresponding to the air temperature and humidity in the time series dataset; determine an evaluation result based on the icing thickness prediction data and the actual icing thickness data, where the evaluation result includes the mean square error, mean absolute error, and mean error magnitude between the icing thickness prediction data and the actual icing thickness data; when the mean square error, mean absolute error, and mean error magnitude are all greater than the corresponding thresholds, it indicates that the icing thickness target prediction model meets the prediction accuracy requirements; when any one of the mean square error, mean absolute error, and mean error magnitude is less than or equal to the corresponding threshold, the hyperparameters of the icing thickness target prediction model are adjusted to obtain a new icing thickness target prediction model.

[0237] Each module in the above-mentioned icing thickness prediction device can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in the processor in the computer device in hardware form or be independent of it, or be stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0238] In an exemplary embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 9 shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store icing thickness prediction and actual data. The input / output interface of the computer device is used for the processor to exchange information with external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements an icing thickness prediction method.

[0239] Those skilled in the art can understand that Figure 9 the structure shown in

[0240] In one embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the foregoing method embodiments are implemented.

[0241] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the foregoing method embodiments are implemented.

[0242] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the foregoing method embodiments are implemented.

[0243] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, artificial intelligence (AI) processors, etc., without limitation.

[0244] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope recorded in the present application.

[0245] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A method for predicting ice coating thickness, characterized in that The method includes: Obtaining a time series data set, where the time series data set includes temperature and humidity; Inputting the temperature and the humidity into a target icing thickness prediction model, and obtaining icing thickness prediction data output by the target icing thickness prediction model; the target icing thickness prediction model is generated by training a model including a convolutional neural network, a long short-term memory network, and an attention mechanism based on a training set.

2. The method according to claim 1, wherein The training method of the target icing thickness prediction model includes: Obtaining a time series data sample set; Processing the time series data sample set based on a convolutional neural network unit, a long short-term memory network unit, and an attention mechanism unit to obtain a key feature vector; Training an initial icing thickness prediction model using the key feature vector to obtain the target icing thickness prediction model.

3. The method according to claim 2, characterized in that, The processing of the time series data sample set based on a convolutional neural network unit, a long short-term memory network unit, and an attention mechanism unit to obtain a key feature vector includes: Processing the time series data sample set through a convolutional neural network unit to obtain the spatial features of the time series data sample set; Extracting the spatio-temporal features of the time series data sample set through a long short-term memory network unit; Processing the spatio-temporal features through the attention mechanism unit to obtain the key feature vector.

4. The method according to claim 3, wherein The convolutional neural network unit includes a convolutional layer and a pooling layer; the processing of the time series data sample set through the convolutional neural network unit to obtain the spatial features of the time series data sample set includes: Performing convolutional processing on the time series data sample set through the convolutional layer to obtain a data sequence; Performing pooling processing on the data sequence through the pooling layer to obtain the spatial features.

5. The method according to claim 3, wherein The processing of the spatio-temporal features through the attention mechanism unit to obtain the key feature vector includes: Determining an attention score according to the obtained first weight matrix and first bias matrix; the first weight matrix is the weight matrix from the input layer to the hidden layer of the attention mechanism unit; the first bias matrix is the bias matrix of the attention mechanism unit; Determining the weight corresponding to the spatio-temporal features according to the attention score; Determining the key feature vector according to the spatio-temporal features and the corresponding weights.

6. The method according to claim 1, characterized in that, After obtaining the icing thickness prediction data output by the target icing thickness prediction model, the method includes: Obtaining the actual icing thickness data corresponding to the temperature and humidity in the time series data set; Determining an evaluation result based on the icing thickness prediction data and the actual icing thickness data, where the evaluation result includes the mean squared error, mean absolute error, and mean error magnitude between the icing thickness prediction data and the actual icing thickness data; When the mean squared error, the mean absolute error, and the mean error magnitude are all greater than the corresponding thresholds, it indicates that the target icing thickness prediction model meets the prediction accuracy requirements; When any one of the mean squared error, the mean absolute error, and the mean error magnitude is less than or equal to the corresponding threshold, the hyperparameters of the icing thickness target prediction model are adjusted to obtain a new icing thickness target prediction model.

7. An icing thickness prediction device, characterized in that, The device includes: an acquisition module, configured to acquire a time series data set, where the time series data set includes temperature and humidity; a processing module, configured to input the temperature and the humidity into an icing thickness target prediction model, and acquire icing thickness prediction data output by the icing thickness target prediction model; the icing thickness target prediction model is generated by training a model including a convolutional neural network, a long short-term memory network, and an attention mechanism based on a training set.

8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images