A power transmission tower swaying prediction method and system based on GBDT-LSTM

Abstract

The application provides a power transmission tower swaying prediction method and system based on GBDT-LSTM. The application obtains power transmission tower environment temperatures, power transmission tower environment humidities, power transmission tower environment wind speeds and real power transmission tower environment swaying displacements at multiple historical moments; the power transmission tower environment temperatures, the power transmission tower environment humidities, the power transmission tower environment wind speeds and the real power transmission tower environment swaying displacements at the multiple historical moments are input into a tower swaying prediction network training module, which is used for combining the real power transmission tower environment swaying displacements at each historical moment to construct a loss function model, and optimizing and training the loss function model through a gradient descent algorithm to obtain an optimized power transmission tower swaying prediction network; and the optimized power transmission tower swaying prediction network is used for prediction to obtain real-time predicted power transmission tower environment swaying displacements. The application improves the accuracy and efficiency of power transmission tower health monitoring and provides a scientific basis for safe operation of a power system.

Description

Technical Field

[0001] This invention belongs to the field of transmission tower sway prediction technology, and particularly relates to a method and system for predicting transmission tower sway based on GBDT-LSTM. Background Technology

[0002] In the current research field of transmission tower sway prediction, although some studies have been conducted based on physical models and simple machine learning methods, these methods often struggle to accurately capture the dynamic characteristics of transmission tower sway under complex environmental conditions. Especially under extreme weather conditions, such as strong winds and thunderstorms, existing technologies such as drone monitoring and infrared camera monitoring often fail to function effectively due to obstructed visibility or equipment malfunction. This results in the inability to obtain accurate transmission tower status information at critical moments, thereby increasing safety risks.

[0003] To address the aforementioned issues, this patent proposes a transmission tower sway prediction model based on Gradient Boosting Decision Trees (GBDT) and Long Short-Term Memory (LSTM) networks, comprehensively considering both internal and external factors contributing to tower sway. Gradient Boosting Decision Trees (GBDT) is an ensemble learning algorithm that sequentially trains decision trees, with each tree correcting the residual (the difference between the actual and predicted values) of the previous tree, thereby progressively improving model performance. GBDT is a powerful and widely used supervised learning algorithm in practice. Long Short-Term Memory (LSTM) is a variant of Recurrent Neural Networks (RNNs) designed to better capture and remember long-term dependencies in time-series data. LSTM effectively solves the problems of gradient vanishing and gradient explosion in traditional RNNs when processing long sequences by introducing special internal memory units and gating mechanisms. GBDT's advantages lie in its efficiency in handling nonlinear problems and feature selection, compensating for LSTM's shortcomings on small datasets; LSTM excels at capturing long-term dependencies in time series data. The combination of GBDT and LSTM leverages the high efficiency of GBDT and the deep time series analysis capabilities of LSTM, improving the overall performance and accuracy of the prediction model. This invention further incorporates a dual adaptive feature transformation mechanism, a dynamic information fusion regulator, and an attention mechanism on top of GBDT and LSTM, further enhancing the model's prediction accuracy. The collaborative model of GBDT and LSTM comprehensively analyzes and utilizes both internal and external factors related to transmission tower sway, significantly improving the accuracy and practicality of the model's predictions.

[0004] This patented model uses a weather forecasting API to obtain weather conditions for the next few days and the current wear and tear on the transmission tower, enabling it to accurately predict the future swaying trend of the tower. This predictive capability allows the model not only to issue early warnings but also to provide decision support for the maintenance of the transmission tower, such as scheduling maintenance personnel for early inspections or, if necessary, evacuating people from nearby areas, thus possessing significant social value and economic benefits.

[0005] With the rapid deployment of 5G networks and the advancement of smart city construction, the requirements for the stability and security of infrastructure such as power transmission towers are increasing. The GBDT-LSTM power transmission tower sway prediction model proposed in this patent not only provides a new solution for current power transmission tower sway prediction technology but also lays the foundation for the future development of intelligent monitoring and maintenance technologies for infrastructure. Through continuous optimization and application of this model, it is expected to play a significant role in ensuring infrastructure safety and promoting smart city construction. Summary of the Invention

[0006] To address the problem of transmission tower sway prediction, this invention proposes a method and system for transmission tower sway prediction based on GBDT-LSTM, aiming to improve the accuracy and efficiency of transmission tower health monitoring and provide a scientific basis for the safe operation of the power system.

[0007] The method of this invention is a transmission tower sway prediction method based on GBDT-LSTM, as detailed below:

[0008] Step 1: Obtain the ambient temperature, humidity, wind speed, and actual swaying displacement of the transmission tower at multiple historical moments.

[0009] Step 2: Construct a transmission tower sway prediction network. Input the transmission tower ambient temperature, humidity, and wind speed at each historical moment into the transmission tower sway prediction network to make predictions and obtain the predicted transmission tower ambient sway displacement at each historical moment. Combine the actual transmission tower ambient sway displacement at each historical moment to construct a loss function model. Optimize and train the model using the gradient descent algorithm to obtain the optimized transmission tower sway prediction network.

[0010] Preferably, the transmission tower sway prediction network in step 2 includes:

[0011] Improved gradient boosting decision tree network, improved long short-term memory network, attention mechanism network;

[0012] The improved gradient boosting decision tree network, the improved long short-term memory network, and the attention mechanism network are cascaded in sequence.

[0013] The improved gradient boosting decision tree network is used to input the transmission tower ambient temperature, transmission tower ambient humidity, and transmission tower ambient wind speed at each historical moment. Through multiple decision tree residual modules, the transmission tower structure loss index at each historical moment is obtained and output to the improved long short-term memory network.

[0014] The improved Long Short-Term Memory (LSTM) network is used to input the transmission tower ambient temperature, humidity, wind speed, and structural wear index at each historical moment. By fusing multiple dynamic information, the LSTM unit is adjusted to obtain the fused hidden state at each historical moment, and then output to the attention mechanism network.

[0015] The attention mechanism network is used as input to the fused hidden state sequence of the improved long short-term memory network. By calculating attention weights and performing weighted summaries, the predicted displacement of the transmission tower sway at each historical moment is obtained.

[0016] The improved gradient boosting decision tree network includes:

[0017] It consists of an adaptive feature learning layer, a first decision tree residual module, a second decision tree residual module, ..., and an Mth decision tree residual module;

[0018] The adaptive feature learning layer, the first decision tree residual module, the second decision tree residual module, ..., the Mth decision tree residual module are cascaded in sequence;

[0019] The m-th decision tree residual module is composed of a decision tree and an adaptive residual module cascaded together.

[0020] The adaptive feature learning layer is used to input the transmission tower ambient temperature, transmission tower ambient humidity, and transmission tower ambient wind speed at each historical moment. Through nonlinear transformation and feature interaction processing, a complex feature vector for each historical moment is obtained and input into the first decision tree residual module.

[0021] The first decision tree residual module is used to input the complex feature vector at each historical moment. The decision tree of the first decision tree residual module repeatedly selects the best feature, divides the dataset, and recursively constructs subtrees. The adaptive residual module introduces weight factors and calculates gradients to obtain the transmission tower structure loss index and adaptive residual at each historical moment. Then, the complex feature vector and adaptive residual at each historical moment are output to the second decision tree residual module.

[0022] The m-th decision tree residual module is used to input the adaptive residual output by the (m-1)-th decision tree residual module and the complex feature vector of each historical moment. It sequentially processes the decision tree of the m-th decision tree residual module by repeatedly selecting the best feature, dividing the dataset, and recursively constructing subtrees. The adaptive residual module introduces weight factors and calculates gradients to obtain the transmission tower structure loss index and adaptive residual of each historical moment. Then, the complex feature vector and adaptive residual of each historical moment are output to the (m+1)-th decision tree residual module.

[0023] m∈[1,M];

[0024] The improved long short-term memory network includes:

[0025] It consists of a first dynamic information fusion and adjustment LSTM unit, a second dynamic information fusion and adjustment LSTM unit, ..., and a Kth dynamic information fusion and adjustment LSTM unit;

[0026] The first dynamic information fusion adjustment LSTM unit, the second dynamic information fusion adjustment LSTM unit, ..., the Kth dynamic information fusion adjustment LSTM unit are cascaded in sequence;

[0027] The kth dynamic information fusion regulation LSTM unit is composed of cascaded LSTM units and dynamic information fusion regulators.

[0028] The first dynamic information fusion and adjustment LSTM unit is used to input the transmission tower ambient temperature, humidity, wind speed, and structural loss index at historical time 1. These data are then passed through the input gate, forget gate, and output gate of the LSTM unit, and through the weighted fusion and decision-making of the dynamic information fusion regulator of the first dynamic information fusion and adjustment LSTM unit, to obtain the fused hidden state at historical time 1. The output is fed to the second dynamic information fusion adjustment LSTM unit and attention mechanism;

[0029] The k-th dynamic information fusion adjustment LSTM unit is used to input the transmission tower ambient temperature, transmission tower ambient humidity, transmission tower ambient wind speed at historical time k, and the fusion hidden state of the (k-1)-th dynamic information fusion adjustment LSTM unit at historical time k-1. By sequentially adjusting the input gate, forget gate, and output gate of the LSTM unit through the k-th dynamic information fusion adjustment, and by weighted fusion and decision-making of the dynamic information fusion regulator of the k-th dynamic information fusion adjustment LSTM unit, the fused hidden state at historical time k is obtained. The output is used to adjust the LSTM unit and attention mechanism through dynamic information fusion at the (k+1)th byte.

[0030] k∈[1,K];

[0031] The loss function model described in step 2 is defined as follows:

[0032]

[0033] in, Time point i is the predicted value of the transmission tower sway displacement, y i is the actual value of the transmission tower sway displacement at time point i, and n is the total number of time points. The lower the MSE, the better the prediction effect of the GBDT-LSTM model.

[0034] Step 3: Obtain real-time ambient temperature, humidity, and wind speed of the transmission tower, and input them into the optimized transmission tower sway prediction network for prediction to obtain the real-time predicted transmission tower environmental sway displacement.

[0035] The technical solution of this invention is a transmission tower sway prediction system based on GBDT-LSTM, comprising:

[0036] The data acquisition module is used to acquire the transmission tower ambient temperature, transmission tower ambient humidity, transmission tower ambient wind speed, and actual transmission tower ambient swaying displacement at multiple historical moments.

[0037] The transmission tower sway prediction network training module is used to construct a loss function model by combining the real transmission tower environmental sway displacement at each historical moment, and to obtain the optimized transmission tower sway prediction network through gradient descent algorithm.

[0038] The real-time prediction module is used to acquire real-time environmental temperature, humidity, and wind speed of the transmission tower, and input them into the optimized transmission tower sway prediction network for prediction, so as to obtain the real-time predicted environmental sway displacement of the transmission tower.

[0039] The present invention offers the following advantages: The transmission tower sway prediction method effectively combines the strengths of GBDT and LSTM networks, comprehensively considering both internal and external factors of transmission tower sway. This not only improves the accuracy of transmission tower health status monitoring but also enables dynamic prediction of transmission tower sway, providing strong technical support for the safe operation of power systems. Furthermore, this method possesses strong universality and adaptability, and can be widely applied to transmission tower monitoring under various environments and conditions. Attached Figure Description

[0040] Figure 1 : Flowchart of the method according to an embodiment of the present invention; Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0043] Data on the swaying displacement of transmission towers is collected using existing methods such as GNSS measurement, laser measurement, tilt measurement, video technology measurement, low-frequency vibration displacement sensor measurement, and tiltmeter measurement. The environmental factors, including temperature, humidity, and wind speed, from multiple historical vectors are obtained through the connected weather forecast data interface.

[0044] The following is combined with Figure 1 Specific embodiments of the present invention are described below.

[0045] like Figure 1 As shown, the specific embodiment of the method of the present invention is a transmission tower sway prediction method based on GBDT-LSTM, as detailed below;

[0046] Step 1: Obtain the ambient temperature, humidity, wind speed, and actual swaying displacement of the transmission tower at multiple historical moments.

[0047] Step 2: Construct a transmission tower sway prediction network. Input the transmission tower ambient temperature, humidity, and wind speed at each historical moment into the transmission tower sway prediction network to make predictions and obtain the predicted transmission tower ambient sway displacement at each historical moment. Combine the actual transmission tower ambient sway displacement at each historical moment to construct a loss function model. Optimize and train the model using the gradient descent algorithm to obtain the optimized transmission tower sway prediction network.

[0048] The transmission tower sway prediction network described in step 2 includes:

[0049] Improved gradient boosting decision tree network, improved long short-term memory network, attention mechanism network;

[0050] The improved gradient boosting decision tree network, the improved long short-term memory network, and the attention mechanism network are cascaded in sequence.

[0051] The improved gradient boosting decision tree network is used to input the transmission tower ambient temperature, transmission tower ambient humidity, and transmission tower ambient wind speed at each historical moment. Through multiple decision tree residual modules, the transmission tower structure loss index at each historical moment is obtained and output to the improved long short-term memory network.

[0052] The improved Long Short-Term Memory (LSTM) network is used to input the transmission tower ambient temperature, humidity, wind speed, and structural wear index at each historical moment. By fusing multiple dynamic information, the LSTM unit is adjusted to obtain the fused hidden state at each historical moment, and then output to the attention mechanism network.

[0053] The attention mechanism network is used as input to the fused hidden state sequence of the improved long short-term memory network. By calculating attention weights and performing weighted summaries, the predicted displacement of the transmission tower sway at each historical moment is obtained.

[0054] The improved gradient boosting decision tree network includes:

[0055] It consists of an adaptive feature learning layer, a first decision tree residual module, a second decision tree residual module, ..., and an Mth decision tree residual module;

[0056] The adaptive feature learning layer, the first decision tree residual module, the second decision tree residual module, ..., the Mth decision tree residual module are cascaded in sequence;

[0057] The m-th decision tree residual module is composed of a decision tree and an adaptive residual module cascaded together.

[0058] The adaptive feature learning layer is used to input the transmission tower ambient temperature, transmission tower ambient humidity, and transmission tower ambient wind speed at each historical moment. Through nonlinear transformation and feature interaction processing, a complex feature vector for each historical moment is obtained and input into the first decision tree residual module.

[0059] The first decision tree residual module is used to input the complex feature vector at each historical moment. The decision tree of the first decision tree residual module repeatedly selects the best feature, divides the dataset, and recursively constructs subtrees. The adaptive residual module introduces weight factors and calculates gradients to obtain the transmission tower structure loss index and adaptive residual at each historical moment. Then, the complex feature vector and adaptive residual at each historical moment are output to the second decision tree residual module.

[0060] The m-th decision tree residual module is used to input the adaptive residual output by the (m-1)-th decision tree residual module and the complex feature vector of each historical moment. It sequentially processes the decision tree of the m-th decision tree residual module by repeatedly selecting the best feature, dividing the dataset, and recursively constructing subtrees. The adaptive residual module introduces weight factors and calculates gradients to obtain the transmission tower structure loss index and adaptive residual of each historical moment. Then, the complex feature vector and adaptive residual of each historical moment are output to the (m+1)-th decision tree residual module.

[0061] m∈[1,M];

[0062] The improved long short-term memory network includes:

[0063] It consists of a first dynamic information fusion and adjustment LSTM unit, a second dynamic information fusion and adjustment LSTM unit, ..., and a Kth dynamic information fusion and adjustment LSTM unit;

[0064] The first dynamic information fusion adjustment LSTM unit, the second dynamic information fusion adjustment LSTM unit, ..., the Kth dynamic information fusion adjustment LSTM unit are cascaded in sequence;

[0065] The kth dynamic information fusion regulation LSTM unit is composed of cascaded LSTM units and dynamic information fusion regulators.

[0066] The first dynamic information fusion and adjustment LSTM unit is used to input the transmission tower ambient temperature, humidity, wind speed, and structural loss index at historical time 1. These data are then passed through the input gate, forget gate, and output gate of the LSTM unit, and through the weighted fusion and decision-making of the dynamic information fusion regulator of the first dynamic information fusion and adjustment LSTM unit, to obtain the fused hidden state at historical time 1. The output is fed to the second dynamic information fusion adjustment LSTM unit and attention mechanism;

[0067] The k-th dynamic information fusion adjustment LSTM unit is used to input the transmission tower ambient temperature, transmission tower ambient humidity, transmission tower ambient wind speed at historical time k, and the fusion hidden state of the (k-1)-th dynamic information fusion adjustment LSTM unit at historical time k-1. By sequentially adjusting the input gate, forget gate, and output gate of the LSTM unit through the k-th dynamic information fusion adjustment, and by weighted fusion and decision-making of the dynamic information fusion regulator of the k-th dynamic information fusion adjustment LSTM unit, the fused hidden state at historical time k is obtained. The output is used to adjust the LSTM unit and attention mechanism through dynamic information fusion at the (k+1)th byte.

[0068] k∈[1,K];

[0069] The loss function model described in step 2 is defined as follows:

[0070]

[0071] in, Time point i is the predicted value of the transmission tower sway displacement, y i Here, MSE is the actual value of the transmission tower sway displacement at time point i, and n is the total number of time points. The lower the MSE, the better the prediction effect of the GBDT-LSTM model.

[0072] The data used in this embodiment of the invention to train the transmission tower sway prediction network is as follows:

[0073] Environment variable sequence: E = {e1, e2, ..., e} T}, where each e t An environmental state vector representing a point in time includes the ambient temperature, humidity, and wind speed of the transmission tower.

[0074] Swaying displacement sequence: D={d1,d2,...,d T}, where each d t This represents the observed displacement of the transmission tower at each point in time.

[0075] The improved GBDT model construction

[0076] The improved GBDT model aims to learn from environmental factors such as transmission tower temperature, humidity, and wind speed to obtain a structural wear index for assessing the degree of structural wear on transmission towers. By utilizing these environmental factors, the model can capture the intrinsic factors leading to structural damage to transmission towers, thereby helping to assess the health status of the towers.

[0077] The improved gradient boosting decision tree network includes:

[0078] It consists of an adaptive feature learning layer, a first decision tree residual module, a second decision tree residual module, ..., and an Mth decision tree residual module;

[0079] The adaptive feature learning layer, by combining multiple nonlinear transformations and a feature interaction layer, can effectively learn complex feature representations from environmental factors, providing rich information for the subsequent GBDT model. The adaptive feature learning layer consists of two ReLU activation layers, a feature interaction layer, and a tanh activation layer stacked in sequence.

[0080] The first layer (ReLU activation layer) introduces a nonlinear transformation through a fully connected layer:

[0081] H1 = σ(W1·E + b1)

[0082] Where W1 is the weight matrix, b1 is the bias term, σ is the activation function (ReLU), H1 is the output of the first hidden layer, and E = {e1, e2, ..., e} T} represents the sequence of environmental variables for the transmission tower, including ambient temperature, humidity, and wind speed.

[0083] The second layer (ReLU activation layer) performs a further nonlinear transformation on H1:

[0084] H2=σ(W2H1+b2)

[0085] Where H1 is the output of the first hidden layer, W2 and b2 are the weight matrix and bias term of the second layer, respectively, σ maintains the same activation function (ReLU), and H2 is the output of the second hidden layer.

[0086] To capture higher-order interactions between features, a third layer (feature interaction layer) can be introduced:

[0087] H interact =ReLU(W interact ·(H2⊙H2)+b interact )

[0088] Here, ⊙ represents element-wise multiplication (Hadamard product), used to simulate the interaction between features (the interaction between the ambient temperature, humidity, and wind speed of the transmission tower). W interact and b interact These are the weight matrix and bias terms. H interact H1 is the output of the higher-order features of the third layer (feature interaction layer), H2 is the output of the first hidden layer, and H3 is the output of the second hidden layer.

[0089] Finally, the fourth layer (output layer) transforms the learned high-order features into the final feature representation:

[0090] Y = tanh(W) o ·H interact +b o )

[0091] Here, W o and b o It is the weight matrix and bias term of the output layer, H interact Y is the high-order feature output of the third layer (feature interaction layer), the tanh hyperbolic tangent activation function, which not only introduces nonlinearity but also compresses the output value to the range of [-1,1]. Y is the complex nonlinear feature representation of the AFL layer output.

[0092] Adaptive residual module

[0093] The adaptive residual module adjustment described above is a strategy for dynamically adjusting the residual calculation method, aiming to improve the model's adaptability to specific samples or data distributions, thereby enhancing overall prediction performance. In the GBDT model, residuals typically refer to the difference between the target value and the model's predicted value, used to guide subsequent decision tree learning. By adaptively adjusting these residuals, we can make the model focus more on time points that are difficult to predict or have a high rate of mispredictions. In this invention example, the SDI structural loss index predicted by the GBDT model is a custom-defined index used to assess the degree of internal loss of transmission towers; however, there is no corresponding real-value data to calculate the residuals in real-world production. Therefore, the adaptive residual module uses the residuals between the transmission tower sway displacement predicted by the improved LSTM model and the attention mechanism and the actual displacement, introduces weighting factors, calculates gradients, and further updates the model parameters.

[0094] Suppose that in a certain iteration, the predicted displacement of the improved LSTM model and attention mechanism at time points i is... The actual displacement is y i The traditional GBDT residual calculation method is as follows:

[0095]

[0096] In adaptive residual adjustment, this embodiment of the invention introduces a weighting factor w. i This makes the adjusted residuals as follows:

[0097]

[0098] Weighting factor w i The weights can be determined using various strategies; this embodiment of the invention employs weight adjustment based on gradient information.

[0099]

[0100] Where L is the loss function defined in step 2. The loss function represents the predictions made by the improved LSTM model and attention mechanism. The gradient is denoted by β, and β is the adjustment coefficient. This method uses gradient information to adjust the residuals, reflecting the model's estimate of its potential to improve loss under the current prediction.

[0101] Ensemble learning of the improved GDBT

[0102] The improved GBDT ensemble learning first initializes the decision tree in the first decision tree residual module, using E = {e1, e2, ..., e...} TThe decision tree in the first decision tree residual module is trained using environmental variable sequence data. This decision tree, through repeated selection of the best features, splitting the dataset, and recursively constructing subtrees, obtains the transmission tower structure loss index, which is then input into the subsequent improved LSTM model and attention mechanism to obtain the predicted displacement of the transmission tower sway. The residual between the predicted displacement and the observed displacement of the transmission tower sway is calculated using the adaptive residual module in the first decision tree residual module. Next, train the next decision tree residual module to fit these residuals. The newly trained decision tree module is then concatenated serially with the previous decision tree residual modules. This process is repeated until the maximum number of iterations is reached. All decision tree residual modules are concatenated serially to form the final improved GBDT model, thus obtaining the final structural loss index sequence S = {s1, s2, ..., s...}. T}, where each s t This represents the observed displacement of the transmission tower at each point in time.

[0103] The improved LSTM construction

[0104] Improved long short-term memory networks include:

[0105] It consists of a first dynamic information fusion and adjustment LSTM unit, a second dynamic information fusion and adjustment LSTM unit, ..., and a Kth dynamic information fusion and adjustment LSTM unit;

[0106] The dynamic information fusion adjustment LSTM unit consists of an LSTM unit and a dynamic information fusion adjuster.

[0107] The LSTM unit of the first dynamic information fusion adjustment LSTM unit is used to input the ambient temperature of the power tower, the ambient humidity of the power tower, the ambient wind speed of the power tower, and the structural loss index of the power tower at historical time 1, and output the hidden state h1 and the cell state C1 at the current time step 1.

[0108] The LSTM unit of the t-th dynamic information fusion adjustment LSTM unit is used as input for the historical time t, including the tower ambient temperature, transmission tower ambient humidity, transmission tower ambient wind speed, transmission tower structural loss index, and the fusion hidden state of the LSTM unit of the (t-1)-th dynamic information fusion adjustment LSTM unit. and cell state C t-1 Output the hidden state h at the current time step t. t and cell state C t .

[0109] The LSTM unit for dynamic information fusion regulation includes three gates: a forget gate, an input gate, and an output gate, as well as a cell state update gate. The forget gate, input gate, and output gate are cascaded in series and then connected in parallel with the cell state update gate.

[0110] The first step is the forgetting gate, which determines whether a memory from a previous moment should be forgotten. The formula for the forgetting gate is:

[0111]

[0112] Among them, f t It is the output of the forget gate, σ is the sigmoid function, and W is the output of the forget gate. f It is the weight matrix of the forget gate. It is the fusion hidden state of the previous time step t-1, x t The input x at the current time step t t =[e t ;s t ;d t-1 ], b f It is a bias term.

[0113] The second step involves the input gate determining which information will be updated in the memory cells. It includes a sigmoid activation function that produces an output between 0 and 1, representing the importance of each input. The formula for the input gate is:

[0114]

[0115] Among them, i t It is the output of the input gate, σ is the sigmoid function, and W is the output of the input gate. i It is the weight matrix of the input gate. It is the fusion hidden state of the previous time step, x t The input x at the current time step t t =[e t ;s t ;d t-1 ], b i It is a bias term.

[0116] The third step, after the input gate and the forget gate, is to update the cell state. The cell state update formula is as follows:

[0117]

[0118] Among them, C t f is the cell state at the current time step t. t It is the output of the forget gate, C t-1 This refers to the cell state at the previous time step t-1, i t It is the output of the input gate, tanh is the hyperbolic tangent activation function, and W is the output of the input gate. c It is the weight matrix for updating the cell state. It is the fusion hidden state of the previous time step t-1, x t The input x at the current time step tt =[e t ;s t ;d t-1 ],b c This is a bias term.

[0119] The fourth step is to determine the previous fusion hidden state through the output gate. Which information will be output to the hidden state at the current moment? The formula for the output gate is:

[0120]

[0121] Among them, o t It is the output of the output gate, σ is the sigmoid activation function, and W is the output of the output gate. o It is the weight matrix of the forget gate, x t The input x is the current historical moment t. t =[e t ;s t ;d t-1 ], It is the fusion hidden state from the previous moment, b o This is a bias term.

[0122] Finally, the hidden state at the current time step:

[0123] h t =o t ·tanh(C t )

[0124] Among them, h t It represents the hidden state at the current time step, tanh is the hyperbolic tangent activation function, and C is the hidden state at the current time step. t It represents the cell state at the current time step t.

[0125] Dynamic information fusion regulator.

[0126] The described state information fusion regulator can not only adjust the fusion ratio among the structural loss index from the improved GBDT, environmental variables (transmission tower ambient temperature, transmission tower humidity, transmission tower wind speed), and transmission tower sway displacement, but also dynamically adjust the weights of the improved GBDT and the improved LSTM outputs based on the current context. This design allows the model to flexibly adjust the information fusion strategy according to different inputs and conditions, thereby optimizing prediction performance.

[0127] The dynamic information fusion regulator of the t-th dynamic information fusion regulating LSTM unit is used to input the hidden state h of the LSTM unit of the dynamic information fusion regulating LSTM unit. t Output the hidden state

[0128] The dynamic information fusion regulator consists of a weight layer, a first Dropout layer, a fusion layer, and a second Dropout layer stacked and cascaded in sequence.

[0129] The weight layer is used as input to the hidden state h output by the LSTM unit at each time step. t Calculate three weights: α t (Weights for SDI) and β t (Weights for environmental variables), γ t (Weights for swaying displacement) The weight layer is defined as follows:

[0130] α t ,β t ,γ t =Softmax(W g ·h t +b g )

[0131] Where Softmax is the activation function, h t It is the hidden state of the time step output, W g It is a weight matrix, b g It is a bias term.

[0132] The first Dropout layer is used to input the weight matrix W of the weight layer. g The weight matrix W of the weight layer is randomly selected. g This is achieved by setting a subset of elements to zero. This effectively prevents overfitting of the model, and is defined as follows:

[0133]

[0134] Here, mask is a binary matrix composed of 0s and 1s, where 1s indicate which elements are retained and 0s indicate which are set to zero. ⊙ represents the element-wise multiplication operation (Hadamard product).

[0135] The fusion layer is used to input the weights α obtained from the weight layer. t ,β t γ t , and the hidden state h of LSTM t Combine and output the hidden state. The next dynamic information fusion adjustment LSTM unit;

[0136] The fusion layer is defined as follows:

[0137]

[0138] in, It is a merged hidden state Wc It is the weight matrix. α t ,β t γ t These are the weights for SDI, the weights for environmental variables, and the weights for sway displacement, respectively. t It is the hidden state of the time step output, b c It is a bias term.

[0139] The second Dropout layer is used to input the weight matrix W of the fusion layer. c The weight matrix W of the weight layer is randomly selected. c This is achieved by setting a subset of elements to zero. The definition is as follows:

[0140]

[0141] Here, mask is a binary matrix composed of 0s and 1s, where 1s indicate which elements are retained and 0s indicate which are set to zero. ⊙ represents the element-wise multiplication operation (Hadamard product).

[0142] Attention mechanism

[0143] To significantly improve the model's ability to capture key information, a Long Short-Term Memory (LSTM) network is cascaded in sequence with an attention mechanism. The core idea of ​​the attention mechanism is to assign different weights to different parts of the input sequence based on contextual information at each time step, enabling the model to "pay attention" to more important information. These weights are then used to calculate a weighted average contextual representation, which in turn influences the model's decision.

[0144] The attention mechanism is used to input the fused hidden state sequence of the improved LSTM as follows: in This is the fused hidden state at time step t. The steps of the attention mechanism being serially cascaded onto the modified LSTM are as follows:

[0145] 1. Calculate attention score

[0146] First, for the fusion hidden state at each time step Calculate the attention score e between the vector and a query vector q. t It indicates Importance:

[0147]

[0148] Where f is a function for calculating the attention score, and the example in this invention uses dot product attention, i.e.:

[0149]

[0150] 2. Calculate attention weights

[0151] Next, the attention scores are converted into weights using a softmax function, ensuring that the sum of all weights is 1:

[0152]

[0153] Here α t Represents the fusion hidden state at time step t. Relative importance in context representation.

[0154] 3. Calculate the weighted context representation

[0155] Using the calculated attention weights, a weighted sum is taken over all fused hidden states to obtain a weighted context representation c:

[0156]

[0157] This context vector c captures important information about the entire input sequence.

[0158] 4. Represent in context

[0159] Finally, the context vector c is fused with the hidden state at the current time step. And use it for subsequent forecasting and decision-making:

[0160]

[0161] Where tanh is the hyperbolic tangent activation function, W c It is a weight matrix, b c It is a bias term, s t It is the predicted displacement of the transmission tower sway at time step t.

[0162] Collaborative optimization

[0163] The improved GBDT, the improved LSTM, and the attention mechanism are synergistically optimized to combine their advantages and improve the accuracy of transmission tower sway displacement prediction. First, the improved GBDT processes the input features and outputs the transmission tower loss index (SDI), which serves as an input feature for the improved LSTM. Then, the improved LSTM and the attention mechanism comprehensively consider the GBDT's output SDI and other raw time-series data (transmission tower sway displacement, ambient temperature, humidity, and wind speed) to predict the transmission tower sway displacement. Through an iterative process, the parameters of the improved GBDT, the improved LSTM, and the attention mechanism are continuously adjusted and optimized using the prediction results of the improved LSTM model and the attention mechanism, thereby improving the performance of the three models. This strategy, through multiple rounds of iterative optimization, ensures high accuracy and robustness of the model in predicting transmission tower sway displacement.

[0164] The optimizer used in this embodiment of the invention is AdamW. AdamW (Adam Weight Decay) is a variant of the Adam optimization algorithm, mainly designed to address some issues with weight decay in the Adam algorithm. AdamW modifies the calculation method of weight decay based on Adam to more accurately implement weight decay and prevent over-updating of weights. The main modification of AdamW is that it adds the weight decay term directly to the loss function, instead of handling weight decay in the update step as in the original Adam algorithm. This allows for more accurate control of the impact of weight decay and improves the stability of the optimization algorithm. The formula for AdamW is as follows:

[0165] Adam update rules:

[0166]

[0167]

[0168]

[0169]

[0170]

[0171] Where, θ t It is a parameter vector. β1 and β2 are the gradient of the loss function with respect to the parameters, η is the learning rate, ∈ is a small constant added for numerical stability, and λ is the weight decay coefficient, controlling the strength of weight decay. This is achieved by using the weight decay term λ·θ. tBy adding it directly to the loss function, AdamW can handle weight decay more accurately, avoiding the problem of excessive weight updates that may occur in the original Adam algorithm. This makes AdamW more stable when training neural networks, especially when dealing with large-scale deep learning models.

[0172] Step 3: Obtain real-time ambient temperature, humidity, and wind speed of the transmission tower, and input them into the optimized transmission tower sway prediction network for prediction to obtain the real-time predicted transmission tower environmental sway displacement.

[0173] Table 1: Prediction Results of Transmission Tower Sway Prediction Network for Certain Time Periods

[0174]

[0175] In the time period captured in Table 1, the mean square error (MSE) between the observed displacement and the predicted displacement of the transmission tower sway is 0.088.

[0176] Table 2: Total Mean Square of Error in Transmission Tower Sway Prediction Network

[0177] 34561 0.091

[0178] Table 2 includes all time periods. The mean square error (MSE) between the observed displacement and the predicted displacement of the transmission tower sway is 0.091. Both tables indicate that the transmission tower sway prediction network has a good prediction effect.

[0179] The technical solution of the system implementation of the present invention is a transmission tower sway prediction system based on GBDT-LSTM, comprising:

[0180] The data acquisition module is used to acquire the transmission tower ambient temperature, transmission tower ambient humidity, transmission tower ambient wind speed, and actual transmission tower ambient swaying displacement at multiple historical moments.

[0181] The transmission tower sway prediction network training module is used to construct a loss function model by combining the real transmission tower environmental sway displacement at each historical moment, and to obtain the optimized transmission tower sway prediction network through gradient descent algorithm.

[0182] The real-time prediction module is used to acquire real-time environmental temperature, humidity, and wind speed of the transmission tower, and input them into the optimized transmission tower sway prediction network for prediction, so as to obtain the real-time predicted environmental sway displacement of the transmission tower.

[0183] The data acquisition module, the transmission tower sway prediction network training module, and the real-time prediction module are all deployed on the server.

[0184] In this embodiment of the invention, the GBDT-LSTM model deployed on each transmission tower has the ability to monitor and predict the tower's sway displacement in real time. Over time, natural factors such as rain erosion lead to a gradual increase in the structural loss of the transmission tower, and correspondingly, the tower loss index (SDI) also increases. To cope with this change, each transmission tower is equipped with corresponding sensing devices that can accurately collect the tower's sway displacement data. Simultaneously, this embodiment of the invention obtains key information such as ambient temperature, humidity, and wind speed near the transmission tower by accessing a weather forecast API. To maintain the accuracy of model predictions, this embodiment of the invention adopts a strategy of periodic model updates. By analyzing the latest collected data, the model periodically self-adjusts and optimizes to adapt to environmental and structural changes, ensuring the accuracy and timeliness of the tower sway displacement prediction. This dynamic update mechanism enhances the model's adaptability and robustness, providing strong technical support for the safe operation of transmission towers. In this embodiment of the invention, the GBDT and LSTM models are used to continuously collect future temperature, humidity, and wind speed data via a weather forecast API. Based on known SDI and collected environmental factors, the system predicts transmission tower sway at future points in time. If the sway exceeds a certain threshold, an early warning is issued to users, enabling them to take appropriate measures to prevent the hazards caused by the sway.

[0185] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0186] It should be understood that the above description of the embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art can make substitutions or modifications under the guidance of this invention without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for predicting transmission tower sway based on GBDT-LSTM, characterized in that, include: Step 1: Obtain the ambient temperature, humidity, wind speed, and actual swaying displacement of the transmission tower at multiple historical moments. Step 2: Construct a transmission tower sway prediction network. Input the transmission tower ambient temperature, humidity, and wind speed at each historical moment into the transmission tower sway prediction network to make predictions and obtain the predicted transmission tower ambient sway displacement at each historical moment. Combine the actual transmission tower ambient sway displacement at each historical moment to construct a loss function model. Optimize and train the model using the gradient descent algorithm to obtain the optimized transmission tower sway prediction network. The transmission tower sway prediction network includes: Improved gradient boosting decision tree network, improved long short-term memory network, attention mechanism network; The improved gradient boosting decision tree network, the improved long short-term memory network, and the attention mechanism network are cascaded in sequence. The improved gradient boosting decision tree network is used to input the transmission tower ambient temperature, transmission tower ambient humidity, and transmission tower ambient wind speed at each historical moment. Through multiple decision tree residual modules, the transmission tower structure loss index at each historical moment is obtained and output to the improved long short-term memory network. The improved Long Short-Term Memory (LSTM) network is used to input the transmission tower ambient temperature, humidity, wind speed, and structural wear index at each historical moment. By fusing multiple dynamic information, the LSTM unit is adjusted to obtain the fused hidden state at each historical moment, and then output to the attention mechanism network. The attention mechanism network is used as input to the fused hidden state sequence of the improved long short-term memory network. By calculating attention weights and performing weighted summaries, the predicted displacement of the transmission tower sway at each historical moment is obtained. The improved long short-term memory network includes: It consists of a first dynamic information fusion and adjustment LSTM unit, a second dynamic information fusion and adjustment LSTM unit, ..., and a Kth dynamic information fusion and adjustment LSTM unit; The first dynamic information fusion adjustment LSTM unit, the second dynamic information fusion adjustment LSTM unit, ..., the Kth dynamic information fusion adjustment LSTM unit are cascaded in sequence; The kth dynamic information fusion regulation LSTM unit is composed of cascaded LSTM units and dynamic information fusion regulators. The first dynamic information fusion and adjustment LSTM unit is used to input the transmission tower ambient temperature, humidity, wind speed, and structural loss index at historical time 1. These data are then passed through the input gate, forget gate, and output gate of the LSTM unit, and through the weighted fusion and decision-making of the dynamic information fusion regulator of the first dynamic information fusion and adjustment LSTM unit, to obtain the fused hidden state at historical time 1. The output is fed to the second dynamic information fusion adjustment LSTM unit and attention mechanism; The k-th dynamic information fusion adjustment LSTM unit is used to input the transmission tower ambient temperature, transmission tower ambient humidity, transmission tower ambient wind speed at historical time k, and the fusion hidden state of the (k-1)-th dynamic information fusion adjustment LSTM unit at historical time k-1. The fused hidden state at historical time k is obtained by sequentially adjusting the input gate, forget gate, and output gate of the LSTM unit through the dynamic information fusion adjustment of the k-th LSTM unit, and by the weighted fusion and decision of the dynamic information fusion regulator of the k-th LSTM unit. The output is fed to the (k+1)th dynamic information fusion adjustment LSTM unit and attention mechanism; k∈[1,K] Where K is the total number of LSTM units for dynamic information fusion adjustment; Step 3: Obtain real-time ambient temperature, humidity, and wind speed of the transmission tower, and input them into the optimized transmission tower sway prediction network for prediction to obtain the real-time predicted transmission tower environmental sway displacement.

2. The transmission tower sway prediction method based on GBDT-LSTM according to claim 1, characterized in that: The improved gradient boosting decision tree network include: It consists of an adaptive feature learning layer, a first decision tree residual module, a second decision tree residual module, ..., and an Mth decision tree residual module; The adaptive feature learning layer, the first decision tree residual module, the second decision tree residual module, ..., the Mth decision tree residual module are cascaded in sequence; The m-th decision tree residual module is composed of a decision tree and an adaptive residual module cascaded together.

3. The transmission tower sway prediction method based on GBDT-LSTM according to claim 2, characterized in that: The adaptive feature learning layer is used to input the ambient temperature, humidity, and wind speed of the transmission tower at each historical moment. Through nonlinear transformation and feature interaction processing, a complex feature vector for each historical moment is obtained and input into the first decision tree residual module.

4. The transmission tower sway prediction method based on GBDT-LSTM according to claim 3, characterized in that: The first decision tree residual module is used to input the complex feature vector at each historical moment. The decision tree of the first decision tree residual module repeatedly selects the best feature, divides the dataset, and recursively constructs subtrees. The adaptive residual module introduces weight factors and calculates gradients to obtain the transmission tower structure loss index and adaptive residual at each historical moment. Then, the complex feature vector and adaptive residual at each historical moment are output to the second decision tree residual module. The m-th decision tree residual module is used to input the adaptive residual output by the (m-1)-th decision tree residual module and the complex feature vector of each historical moment. It sequentially processes the decision tree of the m-th decision tree residual module by repeatedly selecting the best feature, dividing the dataset, and recursively constructing subtrees. The adaptive residual module introduces weight factors and calculates gradients to obtain the transmission tower structure loss index and adaptive residual of each historical moment. Then, the complex feature vector and adaptive residual of each historical moment are output to the (m+1)-th decision tree residual module. m∈[1,M].

5. The transmission tower sway prediction method based on GBDT-LSTM according to claim 1, characterized in that: The loss function model described in step 2 is defined as follows: in, Time point i is the predicted value of the transmission tower's sway displacement. is the actual value of the swaying displacement of the transmission tower at time point i, and n is the total number of time points.

6. A transmission tower sway prediction system based on GBDT-LSTM, characterized in that, Used to perform the GBDT-LSTM-based transmission tower sway prediction method as described in any one of claims 1-5.

Images