A real-time monitoring method for power engineering construction environment

Abstract

The present application relates to the technical field of data processing, and especially relates to a real-time monitoring method for power engineering construction environment, which obtains a multi-dimensional historical environment data sequence of a power engineering construction area; according to a device running in an LSTM model, an initial interval of a hidden unit quantity when predicting future multi-dimensional environment data by using the LSTM model is obtained; according to the change characteristics of the multi-dimensional historical environment data sequence and the initial interval of the hidden unit quantity, a standard interval of the hidden unit quantity is obtained; any value in the standard interval is taken as the hidden unit quantity in the LSTM model, the LSTM model is trained by using the multi-dimensional historical environment data sequence, and a training effect score of the LSTM model under any value is obtained; by using the training effect score of the LSTM model under each value in the standard interval, an optimal hidden unit quantity is obtained, an LSTM prediction model is obtained, and the accuracy of prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for real-time monitoring of the construction environment in power engineering. Background Technology

[0002] The construction environment for power engineering is complex and changeable. Construction sites typically face multiple risk factors, such as working at heights, operating outdoors, and transporting equipment. Changes in meteorological environmental parameters, such as wind speed, wind direction, temperature, humidity, and air pressure, can directly affect operational safety and progress. For example, strong winds can cause high-altitude equipment (such as cranes and insulated bucket trucks) to become unstable, posing a high risk of falling objects and swaying. Sudden changes in wind direction may cause overhead conductors or suspended equipment to deviate from the work area, leading to accidental collisions or construction delays. Therefore, real-time monitoring of these key meteorological environmental parameters is an indispensable part of the construction scheduling, safety protection, and emergency response system.

[0003] Because the natural environment of power engineering construction sites (such as plateaus, mountains, and coastal areas) is subject to frequent and uncontrollable changes in weather conditions, which directly affect construction safety, it is necessary to issue early warnings before the risks become apparent, so as to allow time for construction scheduling to react.

[0004] Traditional methods typically utilize LSTM models to predict meteorological environmental data. In this scenario, LSTM models are trained on historical data from a previous period before being used in the next stage of prediction, thus issuing early warnings before risks become apparent. Because meteorological environmental data exhibits significant dynamic fluctuations, LSTM models need a certain capacity to remember and capture data features to accurately predict these trends. In LSTM models, the number of hidden units determines the model's ability to remember and process data. However, this is often set based on human experience, relying too heavily on subjective judgment and lacking data-driven support. Setting the number of hidden units too small results in retaining only very short historical dependencies, making it impossible for the LSTM model to capture these patterns. Setting the number of hidden units too large increases training costs and poses a risk of overfitting.

[0005] Therefore, how to adaptively obtain the optimal number of hidden units in the LSTM model and improve the prediction accuracy of the LSTM model has become an urgent problem to be solved. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method for real-time monitoring of the construction environment of power engineering, in order to solve the problem of how to adaptively obtain the optimal number of hidden units in the LSTM model and improve the accuracy of LSTM model prediction.

[0007] This invention provides a method for real-time monitoring of the construction environment in power engineering, the method comprising the following steps: Obtain multidimensional historical environmental data within a preset historical time period in the power engineering construction area to obtain a multidimensional historical environmental data sequence. Based on the device deployed and running the LSTM model, obtain the maximum value of the number of LSTM parameters in the LSTM model, and based on the maximum value, obtain the initial range of the number of hidden units when using the LSTM model to predict future multidimensional environmental data. Based on the variation characteristics of the multidimensional historical environment data sequence, the complexity score of the multidimensional historical environment data sequence is obtained, and based on the complexity score and the initial interval, the standard interval of the number of hidden units is obtained. For any value within the specified range, the specified value is used as the number of hidden units in the LSTM model. The LSTM model is trained using the multidimensional historical environment data sequence according to a preset training round, and the training effect score of the LSTM model under the specified value is obtained. The training performance score of the LSTM model is obtained for each value within the specified range. Based on the training performance score of the LSTM model for each value within the specified range, the optimal number of hidden units is obtained. The optimal number of hidden units is used as the number of hidden units in the LSTM model to obtain an LSTM prediction model for predicting future multidimensional environmental data, which is used for real-time monitoring of the construction environment in the power engineering construction area.

[0008] Preferably, the step of obtaining the maximum number of LSTM parameters in the LSTM model for the device deployed and running according to the LSTM model includes: The device on which the LSTM model is deployed and run is designated as the target device. The available memory of the target device is obtained, and the available memory of the target device is calculated as a preset multiple to obtain the maximum number of LSTM parameters in the LSTM model.

[0009] Preferably, obtaining the initial interval for the number of hidden units when using the LSTM model to predict future multidimensional environmental data based on the maximum value includes: Using the maximum value, the formula for calculating the number of LSTM parameters is reversed to obtain the maximum number of hidden units. The number of dimensions of the multidimensional historical environmental data is then obtained. The number of dimensions and the maximum number of hidden units are used to form the initial interval for the number of hidden units when using the LSTM model to predict future multidimensional environmental data.

[0010] Preferably, obtaining the complexity score of the multidimensional historical environmental data sequence based on its variation characteristics includes: PCA principal component analysis is performed on the multidimensional historical environmental data sequence to obtain at least one principal component and its corresponding variance contribution rate. A scree plot is constructed based on each principal component and its corresponding variance contribution rate. The horizontal axis of the scree plot is the number of principal components, and the vertical axis is the variance contribution rate. Obtain the inflection point of the scree plot. In the scree plot, mark the abscissa of the inflection point as the number of principal components. Use the ratio of the number of principal components to the number of dimensions as the independent variable of the sigmoid function to obtain the first complexity score of the multidimensional historical environment data sequence. The multidimensional historical environment data sequence is divided into a preset number of one-dimensional historical environment data sequences according to the dimensions of the multidimensional historical environment data. The information entropy of each one-dimensional historical environment data sequence is obtained. The information entropy of each one-dimensional historical environment data sequence is used as the independent variable of the sigmoid function, and the mean value of the sigmoid function is obtained, which is denoted as the second complexity score of the multidimensional historical environment data sequence. The complexity score of the multidimensional historical environment data sequence is obtained based on the average between the first complexity score and the second complexity score.

[0011] Preferably, obtaining the standardized interval for the number of hidden units based on the complexity score and the initial interval includes: Obtain the difference between the maximum number of hidden units and the number of dimensions to get the span value of the initial interval, calculate the product between the span value and the complexity score, and obtain the sum of the product and the number of dimensions to get the number of normalized hidden units; The fluctuation range of the specified number of hidden units is set as the specified range for the number of hidden units.

[0012] Preferably, the step of training the LSTM model using the multidimensional historical environment data sequence according to a preset training round, and obtaining a training performance score of the LSTM model for any given value, includes: According to a preset ratio, the multidimensional historical environment data sequence is divided into a multidimensional training data sequence and a multidimensional verification data sequence; For any training epoch, the multidimensional training data sequence is input into the LSTM model for training to obtain the adjusted LSTM model for that training epoch. The average error generated by the LSTM model during training is obtained and recorded as the training error. The multidimensional validation data sequence is input into the adjusted LSTM model for validation to obtain the average error generated by the adjusted LSTM model during validation and recorded as the validation error. The mean between the training error and the validation error is calculated to obtain the overall error value for that training epoch. Obtain the training error, validation error, and total error value for each training round. Based on the training error, validation error, and total error value for each training round, obtain the training performance score of the LSTM model for any given value.

[0013] Preferably, obtaining the training performance score of the LSTM model for any given value based on the training error, validation error, and overall error value of each training epoch includes: The total error values ​​of each training round are combined into a total error value sequence. The total error value sequence is fitted to obtain a fitted curve. The inflection point of the fitted curve is obtained and recorded as the total error inflection point. The training round corresponding to the total error inflection point is recorded as the training convergence round. Substitute the total error value of the last training round into the exponential function with the training convergence round as the base to obtain the exponential function result. Use the exponential function result as the independent variable of the sigmoid function to obtain the corresponding sigmoid function value. Obtain the difference between the constant 1 and the sigmoid function value to obtain the first training effect score. Obtain the absolute value of the difference between the training error and the validation error in each training round, and obtain the cumulative value of the absolute value of the difference. Use the cumulative value of the absolute value of the difference as the independent variable of the sigmoid function to obtain the total difference result. Obtain the difference between the constant 1 and the total difference result to obtain the second training effect score. The training performance score of the LSTM model for any given value is obtained by averaging the first training performance score and the second training performance score.

[0014] Preferably, obtaining the optimal number of hidden units based on the training performance score of the LSTM model for each value within the specified interval includes: The optimal number of hidden units is determined by taking the number of hidden units corresponding to the maximum training performance score within the specified range.

[0015] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows: This invention acquires multidimensional historical environmental data within a preset historical period in a power engineering construction area, obtaining a multidimensional historical environmental data sequence. Based on the equipment deployed and running the LSTM model, it obtains the maximum value of the number of LSTM parameters in the LSTM model. Based on this maximum value, it obtains an initial interval for the number of hidden units when using the LSTM model to predict future multidimensional environmental data. Based on the variation characteristics of the multidimensional historical environmental data sequence, it obtains a complexity score for the multidimensional historical environmental data sequence. Based on the complexity score and the initial interval, it obtains a standardized interval for the number of hidden units. For any value within the standardized interval, it sets the number of hidden units... The value is used as the number of hidden units in the LSTM model. According to the preset training epochs, the LSTM model is trained using the multidimensional historical environmental data sequence to obtain the training effect score of the LSTM model under any given value. The training effect score of the LSTM model under each value within the specified interval is obtained. Based on the training effect score of the LSTM model under each value within the specified interval, the optimal number of hidden units is obtained. The optimal number of hidden units is used as the number of hidden units in the LSTM model to obtain an LSTM prediction model for predicting future multidimensional environmental data, which is used for real-time monitoring of the construction environment in the power engineering construction area. Specifically, based on the equipment on which the LSTM model is deployed, an initial range of hidden unit numbers is obtained when using the LSTM model to predict future multidimensional environmental data. This ensures that the LSTM model has good response speed and prediction accuracy under limited computing resources. Then, based on the complexity score and the initial range, a standardized range of hidden unit numbers is obtained, matching the number of hidden units with the memory capacity of the LSTM model. This further improves the generalization ability and training efficiency of the LSTM model while ensuring prediction accuracy. Finally, based on the training effect score of the LSTM model for each value within the standardized range, the optimal number of hidden units is obtained, making the LSTM prediction model more compatible with historical environmental data and improving prediction accuracy. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for real-time monitoring of the construction environment of power engineering provided in Embodiment 1 of the present invention. Detailed Implementation

[0018] Embodiments of this disclosure are described in detail below, with examples of these embodiments illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting it.

[0019] It should be noted that the terms "first," "second," etc., used in this disclosure and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0020] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0021] See Figure 1 This is a flowchart of a real-time monitoring method for the construction environment of power engineering provided in Embodiment 1 of the present invention, as follows: Figure 1 As shown, the method may include: Step S101: Obtain multi-dimensional historical environmental data within a preset historical time period in the power engineering construction area to obtain a multi-dimensional historical environmental data sequence.

[0022] The construction environment for power engineering projects is complex and variable. Changes in meteorological environmental parameters such as wind speed, wind direction, temperature, humidity, and air pressure directly impact operational safety and progress. Therefore, real-time monitoring of these key meteorological environmental parameters is an indispensable component of construction scheduling, safety protection, and emergency response systems. Because meteorological conditions at power engineering construction sites are frequently and uncontrollably changing, directly affecting construction safety, it is necessary to issue early warnings before risks become apparent, allowing sufficient reaction time for construction scheduling.

[0023] Traditional methods typically utilize LSTM models to predict meteorological environmental data. In this scenario, LSTM models are trained on historical data from a previous period before being used in the next stage of prediction, thus issuing early warnings before risks become apparent. Because meteorological environmental data exhibits significant dynamic fluctuations, LSTM models need a certain capacity to remember and capture data features to accurately predict these trends. In LSTM models, the number of hidden units determines the model's ability to remember and process data. However, this is often set based on human experience, relying too heavily on subjective judgment and lacking data-driven support. Setting the number of hidden units too small results in retaining only very short historical dependencies, making it impossible for the LSTM model to capture these patterns. Setting the number of hidden units too large increases training costs and poses a risk of overfitting.

[0024] Therefore, in this embodiment, firstly, a multidimensional historical environmental data sequence of the power engineering construction area is obtained. Based on the equipment deployed and running the LSTM model, an initial interval for the number of hidden units when using the LSTM model to predict future multidimensional environmental data is obtained, ensuring that the LSTM model has good response speed and prediction accuracy under limited computing resources. Then, based on the variation characteristics of the multidimensional historical environmental data sequence and the initial interval for the number of hidden units, a standardized interval for the number of hidden units is obtained, so that the number of hidden units matches the memory capacity of the LSTM model, thereby further improving the generalization ability and training efficiency of the LSTM model while ensuring prediction accuracy. Finally, the training effect score of the LSTM model for each value within the standardized interval is obtained, thereby obtaining the optimal number of hidden units, making the LSTM prediction model more compatible with historical environmental data and improving the prediction accuracy of the LSTM prediction model.

[0025] In this embodiment, the multidimensional historical environmental data sequence includes wind speed, wind direction, and humidity. There are no restrictions here, and the data can be set according to the specific implementation scenario. The data can be obtained through local weather stations, ultrasonic anemometers, and humidity collectors. The wind speed, wind direction, and humidity within a preset historical period, i.e., a quarter, are obtained at a sampling frequency of once per second. There are no restrictions here, and the data can be set according to the specific implementation scenario. In order to ensure that the data input into the LSTM model has good continuity, stability, and standardization, the collected multidimensional historical environmental data needs to be preprocessed before being assembled into a multidimensional historical environmental data sequence.

[0026] The preprocessing steps are as follows: (1) Fill in the missing data values ​​caused by sensor anomalies, communication delays, etc. using linear interpolation; (2) Eliminate local extreme value jitter by median filtering; (3) If the collection frequency of historical environmental data in each dimension of the multidimensional historical environmental data is inconsistent, align all historical environmental data according to a unified time frequency; (4) Convert wind direction data into sine and cosine values, such as wind direction angle. After conversion, it becomes ( (5) Normalize historical environmental data under different dimensions to avoid imbalance in LSTM model training caused by different units; among them, linear interpolation and median filtering are existing technologies, which will not be elaborated here.

[0027] Step S102: Based on the device where the LSTM model is deployed and running, obtain the maximum value of the number of LSTM parameters in the LSTM model, and based on the maximum value, obtain the initial range of the number of hidden units when using the LSTM model to predict future multidimensional environmental data.

[0028] Since edge computing devices or lightweight models are often deployed locally in the construction environment of power engineering projects, the complexity of the LSTM model must be strictly controlled during the design phase to ensure that the LSTM model has good response speed and prediction accuracy under limited computing resources.

[0029] Since the number of hidden units in an LSTM model directly determines the number of parameters and computational cost of the LSTM model, and there is no unified method for obtaining or quantifying the number of hidden units, in order to avoid blindly searching among a massive number of units, this embodiment uses the upper and lower limits of the number of hidden units allowed by the LSTM model in this implementation scenario as the initial range of the number of hidden units when using the LSTM model to predict future multidimensional environmental data.

[0030] The upper limit of the number of LSTM parameters allowed by the LSTM model is denoted as Pmax. Pmax is mainly limited by the computing power (such as memory, video memory, etc.) of the device on which the LSTM model is deployed and run, and the requirements of the deployment scenario. For example, the relationship between the deployment configuration and Pmax in a common power engineering construction environment (based on float32 parameters, each parameter occupies 4 bytes): Intel-i5 device (8G memory), available memory / video memory is 1000-2000MB, Pmax is 200000-300000; Android device (4G memory), available memory / video memory is 256-512MB, Pmax is 80000-120000. Therefore, in this embodiment, the device on which the LSTM model is deployed and run is denoted as the target device. The available memory of the target device is obtained, and one-quarter of the available memory of the target device is calculated as the maximum number of LSTM parameters in the LSTM model. One-quarter is a preset multiple of the available memory of the target device set in this embodiment, which is not limited here and can be set according to the specific implementation scenario.

[0031] Furthermore, using the maximum value, the formula for calculating the number of LSTM parameters is inversely operated to obtain the maximum number of hidden units, which serves as the upper limit of the number of hidden units allowed by the LSTM model in this implementation scenario, denoted as . The formula for calculating the number of LSTM parameters is: Where P is the number of LSTM parameters, h is the number of hidden units, and D is the number of dimensions of the multidimensional historical environment data. The formula for calculating the number of LSTM parameters is existing technology and will not be elaborated here.

[0032] Since the design principle of the LSTM model is to form a more complex temporal behavior representation space based on multiple time steps than a single step, a higher dimension is needed to represent the hidden state. That is, the number of hidden units is usually greater than the dimension of the input data. Therefore, the number of dimensions of the multi-dimensional historical environment data is used as the lower limit of the number of hidden units allowed by the LSTM model in this implementation scenario, denoted as . .

[0033] The number of dimensions With the maximum number of hidden units The initial interval for the number of hidden units when using an LSTM model to predict future multidimensional environmental data is formed. ].

[0034] Thus, we obtain the initial range of the number of hidden units when using the LSTM model to predict future multidimensional environmental data.

[0035] Step S103: Based on the change characteristics of the multidimensional historical environment data sequence, obtain the complexity score of the multidimensional historical environment data sequence, and based on the complexity score and the initial interval, obtain the standard interval of the number of hidden units.

[0036] In LSTM models, the number of hidden units represents the model's capacity or memory ability, indicating how much information the LSTM model can retain at each time step. However, the specific amount of information the LSTM model needs to retain is highly positively correlated with the complexity of the input data. That is, the more complex the input data, the more hidden units the LSTM model needs to retain information and capture patterns; simpler input data requires fewer hidden units to prevent overfitting. Therefore, the number of hidden units should be set based on the complexity of the input data upon which the LSTM model relies for training, matching the number of hidden units with the LSTM model's memory ability. This ensures that the generalization ability and training efficiency of the LSTM model are further improved while maintaining prediction accuracy. Therefore, it is necessary to obtain a complexity score for the multidimensional historical environment data sequence based on its changing characteristics to evaluate the complexity of the input data upon which the LSTM model relies for training.

[0037] The method for obtaining the complexity score of a multidimensional historical environmental data sequence based on its variation characteristics is as follows: In real-time monitoring of the construction environment of power engineering, the input data of the LSTM model includes multiple meteorological environmental parameters (wind speed, wind direction, humidity). There may be strong correlations or information redundancy among these multiple meteorological environmental parameters. Therefore, in order to clearly know the truly effective "number of information dimensions" in the multidimensional historical environmental data, that is, how many dimensions of information the LSTM model needs to remember during training to grasp the main trends, PCA principal component analysis is performed on the multidimensional historical environmental data sequence to obtain at least one principal component and its corresponding variance contribution rate. A scree plot is constructed based on each principal component and its corresponding variance contribution rate. The horizontal axis of the scree plot is the number of principal components, and the vertical axis is the variance contribution rate. PCA principal component analysis and scree plot are existing technologies and will not be elaborated here. The Kneedl algorithm is used to obtain the inflection point of the scree plot. The Kneedl algorithm is an existing technology and will not be described in detail here. In the scree plot, the inflection point is the point where the curve changes from a rapid descent to a slow descent. The principal components corresponding to the points before the inflection point usually explain most of the variance information. The explanatory power of the principal components after the inflection point tends to be smooth and the contribution decreases. Therefore, all principal components before the inflection point are the main information carriers. The x-coordinate of the inflection point is marked as the number of principal components. The ratio of the number of principal components to the number of dimensions is used as the independent variable of the sigmoid function to obtain the first complexity score of the multidimensional historical environment data sequence. To gain a clearer understanding of the complexity of multidimensional historical environment data, it is necessary to quantify the changes in historical environment data under each dimension. Information entropy can measure the distribution and fluctuation of data values. Therefore, the multidimensional historical environment data sequence is divided into a predetermined number of one-dimensional historical environment data sequences according to the dimensions of the multidimensional historical environment data. The information entropy of each one-dimensional historical environment data sequence is obtained, and the information entropy of each one-dimensional historical environment data sequence is used as the independent variable of the sigmoid function. The mean value of the sigmoid function is obtained and is denoted as the second complexity score of the multidimensional historical environment data sequence. The complexity score of the multidimensional historical environment data sequence is obtained based on the average between the first complexity score and the second complexity score.

[0038] In one embodiment, the formula for calculating the complexity score of the multidimensional historical environmental data sequence is as follows:

[0039] Where S is the complexity score of the multidimensional historical environmental data sequence; K is the number of principal components; and D is the number of dimensions. Let be the information entropy of the i-th one-dimensional historical environment data sequence; The sigmoid function, also known as the sigmoid function or sigmoid growth curve, maps the output to (0, 1).

[0040] It should be noted that, The first complexity score reflects the information density of multidimensional historical environmental data. The larger the K value, the more the multidimensional historical environmental data is distributed across multiple principal components, with less redundancy between dimensions, more dispersed information, higher information density, and more complex data. The larger it is, the larger S will be; The second complexity score reflects the overall volatility of data primarily based on historical environments. The larger the value of S, the more drastic and irregular the changes in the multidimensional historical environmental data are, indicating a chaotic distribution. The more complex the data, the larger the value of S.

[0041] To ensure the number of hidden units matches the complexity of the input data, the complexity scores of the obtained multidimensional historical environmental data sequences are mapped to an initial interval for the number of hidden units, making the LSTM model structure adaptable to the complexity of the input data. Furthermore, in actual construction environments, the complexity of meteorological environmental data typically exhibits continuous variation; for example, wind speed changes may gradually shift from stable and regular to drastic and abrupt changes, rather than abruptly. Linear interpolation naturally supports this continuous mapping logic, smoothly and reasonably converting the complexity scores into the corresponding number of standard hidden units. Therefore, the linear interpolation method is used to map the complexity scores to an initial interval to obtain the standard number of hidden units.

[0042] The method for mapping the complexity score to the initial interval using linear interpolation to obtain the number of canonical hidden units is as follows: Obtain the difference between the maximum number of hidden units and the number of dimensions to get the span value of the initial interval, calculate the product between the span value and the complexity score, and obtain the sum of the product and the number of dimensions to get the number of normalized hidden units; In one embodiment, the formula for calculating the number of standard hidden units is:

[0043] in, To standardize the number of hidden units; S is the complexity score of the multidimensional historical environment data sequence; This represents the maximum number of hidden units. The number of dimensions is used. The formula for calculating the number of hidden units is based on the application of linear interpolation, which is an existing technology and will not be elaborated upon here.

[0044] The canonical number of hidden units reflects the memory capacity required by the LSTM model to understand and recognize the input data given the complexity score; in other words, how many hidden units are needed. Therefore, the canonical number of hidden units strongly indicates the optimal number of hidden units, which lies within a range close to the canonical number.

[0045] Therefore, the fluctuation range of the number of hidden units can be used as the standard range for the number of hidden units. Within the standard range, the value of the number of hidden units for each value is added to the LSTM model for training, thereby obtaining the optimal number of hidden units. However, setting the fluctuation range of the number of hidden units too large may lead to a broad evaluation scope, increased cost, and higher expenses; setting it too small may result in missing the optimal number of hidden units. Therefore, in this embodiment, the fluctuation range of the number of hidden units is set to […]. This avoids the inefficiency of blindly searching within the entire initial interval. There are no restrictions here, and it can be set according to the specific implementation scenario.

[0046] Thus, the standard range for the number of hidden units is obtained.

[0047] Step S104: For any value within the specified range, take the specified value as the number of hidden units in the LSTM model, train the LSTM model using the multidimensional historical environment data sequence according to the preset training rounds, and obtain the training effect score of the LSTM model under the specified value.

[0048] Since the final performance of the LSTM model is primarily reflected by its ability to learn data features and its accuracy in predicting data, after obtaining a standardized range for the number of hidden units, any value within this range is used as the number of hidden units in the LSTM model. The LSTM model is trained using multidimensional historical environmental data sequences according to a preset training epoch, and the training performance score for any given value is obtained. This score is used to evaluate the learning ability and prediction performance of the LSTM model at any given value. Finally, based on the training performance score of the LSTM model for each value within the standardized range, the optimal number of hidden units is determined. In this embodiment, the preset training epoch is set to 50, but this is not a limitation and can be set according to the specific implementation scenario.

[0049] The method for training the LSTM model using multidimensional historical environmental data sequences according to a preset training round and obtaining the training performance score of the LSTM model for any given value is as follows: According to a preset ratio of 8:2, the multidimensional historical environment data sequence is divided into a multidimensional training data sequence and a multidimensional verification data sequence. The time corresponding to the multidimensional training data sequence must be earlier than the time corresponding to the multidimensional verification data sequence. There is no restriction here, and it can be set according to the specific implementation scenario. For any training epoch, each training epoch includes a training process and a validation process. The training process involves inputting the multidimensional training data sequence into the LSTM model to train it, thereby adjusting the parameters of the LSTM model and obtaining the adjusted LSTM model for that training epoch. The average error generated by the LSTM model during the training process is obtained and denoted as the training error. The validation process involves inputting the multidimensional validation data sequence into the adjusted LSTM model for that training epoch to validate it, thereby evaluating the performance of the adjusted LSTM model in the multidimensional validation data sequence. The average error generated by the adjusted LSTM model during the validation process is obtained and denoted as the validation error. The mean between the training error and the validation error is calculated to obtain the overall error value for that training epoch. The training error, validation error, and overall error value for each training epoch are obtained. The overall error values ​​for each training epoch are combined into an overall error value sequence. The overall error value sequence is fitted to obtain a fitted curve. The inflection point of the fitted curve is obtained using the Kneedl algorithm and is denoted as the overall error inflection point. The inflection point is the point where the fitted curve begins to change slowly. The training epoch corresponding to the overall error inflection point is denoted as the training convergence epoch u, indicating that the LSTM model has basically mastered the data features and the overall loss begins to stabilize in the u-th training epoch. The Kneedl algorithm is an existing technology and will not be described in detail here. Substitute the total error value of the last training round into the exponential function with the training convergence round as the base to obtain the exponential function result. Use the exponential function result as the independent variable of the sigmoid function to obtain the corresponding sigmoid function value. Obtain the difference between the constant 1 and the sigmoid function value to obtain the first training effect score. Obtain the absolute value of the difference between the training error and the validation error in each training round, and obtain the cumulative value of the absolute value of the difference. Use the cumulative value of the absolute value of the difference as the independent variable of the sigmoid function to obtain the total difference result. Obtain the difference between the constant 1 and the total difference result to obtain the second training effect score. The training performance score of the LSTM model for any given value is obtained by averaging the first training performance score and the second training performance score.

[0050] In one implementation, taking the j-th value within the normalized interval as an example, the formula for calculating the training performance score of the LSTM model for the j-th value is as follows:

[0051] in, The training performance score of the LSTM model for the j-th value; u is the training convergence epoch; This represents the total error value for the m-th training round (i.e., the last training round); Let be the training error in the c-th training round; Let be the validation error in the c-th training epoch; m be the number of training epochs. The Sigmoid function, also known as the S-shaped function or S-shaped growth curve, maps the output to (0, 1). It is the absolute value symbol.

[0052] It should be noted that, The first training performance score reflects the learning ability of the LSTM model during training. Specifically, the LSTM model converges faster across multiple data metrics. In power engineering construction sites, the available computing resources for edge devices or terminals have a definite upper limit. Therefore, under fixed and limited computing resources, faster convergence indicates higher resource utilization and better model performance. A smaller u indicates faster convergence and better learning ability of the LSTM model. The larger it is, the better. The larger it is; To score the second training effect, The smaller the value, the better the prediction stability of the current LSTM model and the more stable the training process. The larger it is, the better. The larger it is.

[0053] Thus, the training performance score of the LSTM model for any given value is obtained.

[0054] Step S105: Obtain the training performance score of the LSTM model for each value within the specified interval. Based on the training performance score of the LSTM model for each value within the specified interval, obtain the optimal number of hidden units. Use the optimal number of hidden units as the number of hidden units in the LSTM model to obtain an LSTM prediction model for predicting future multidimensional environmental data, which is used for real-time monitoring of the construction environment in the power engineering construction area.

[0055] Following the method for obtaining the training performance score of the LSTM model under any of the above values, the training performance score of the LSTM model under each value within the standard interval is obtained. The larger the training performance score, the better the learning ability and prediction effect of the LSTM model. Therefore, the value of the number of hidden units corresponding to the maximum training performance score within the standard interval is taken as the optimal number of hidden units.

[0056] Furthermore, the optimal number of hidden units is used as the number of hidden units in the LSTM model to obtain an LSTM prediction model for predicting future multi-dimensional environmental data. Real-time collected multi-dimensional historical environmental data is input into the trained LSTM prediction model, which outputs the prediction result for the next time step. The prediction result is compared with a preset engineering safety threshold, and information is pushed to the dispatching system or management personnel based on the comparison result for decision-making, thus completing the real-time monitoring of the power engineering construction environment. The LSTM prediction model obtaining the prediction result for the next time step based on the input data is existing technology and will not be elaborated upon here.

[0057] In summary, this embodiment of the invention acquires multidimensional historical environmental data within a preset historical period in a power engineering construction area to obtain a multidimensional historical environmental data sequence; based on the equipment deployed and running the LSTM model, it obtains the maximum value of the number of LSTM parameters in the LSTM model, and based on the maximum value, obtains an initial interval for the number of hidden units when using the LSTM model to predict future multidimensional environmental data; based on the variation characteristics of the multidimensional historical environmental data sequence, it obtains the complexity score of the multidimensional historical environmental data sequence, and based on the complexity score and the initial interval, it obtains a standardized interval for the number of hidden units; for any value within the standardized interval, it... The number of hidden units in the LSTM model is determined by taking any given value as the number of hidden units. The LSTM model is trained using the multidimensional historical environmental data sequence according to a preset training epoch, and the training performance score of the LSTM model under any given value is obtained. The training performance score of the LSTM model under each value within the specified interval is obtained. Based on the training performance score of the LSTM model under each value within the specified interval, the optimal number of hidden units is obtained. The optimal number of hidden units is taken as the number of hidden units in the LSTM model, resulting in an LSTM prediction model for predicting future multidimensional environmental data, which is used for real-time monitoring of the construction environment in the power engineering construction area. Specifically, based on the equipment on which the LSTM model is deployed, an initial range of hidden unit numbers is obtained when using the LSTM model to predict future multidimensional environmental data. This ensures that the LSTM model has good response speed and prediction accuracy under limited computing resources. Then, based on the complexity score and the initial range, a standardized range of hidden unit numbers is obtained, matching the number of hidden units with the memory capacity of the LSTM model. This further improves the generalization ability and training efficiency of the LSTM model while ensuring prediction accuracy. Finally, based on the training effect score of the LSTM model for each value within the standardized range, the optimal number of hidden units is obtained, making the LSTM prediction model more compatible with historical environmental data and improving prediction accuracy.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for real-time monitoring of the construction environment in power engineering, characterized in that, The method includes: Obtain multidimensional historical environmental data within a preset historical time period in the power engineering construction area to obtain a multidimensional historical environmental data sequence. Based on the device deployed and running the LSTM model, obtain the maximum value of the number of LSTM parameters in the LSTM model, and based on the maximum value, obtain the initial range of the number of hidden units when using the LSTM model to predict future multidimensional environmental data. Based on the variation characteristics of the multidimensional historical environment data sequence, the complexity score of the multidimensional historical environment data sequence is obtained, and based on the complexity score and the initial interval, the standard interval of the number of hidden units is obtained. For any value within the specified range, the specified value is used as the number of hidden units in the LSTM model. The LSTM model is trained using the multidimensional historical environment data sequence according to a preset training round, and the training effect score of the LSTM model under the specified value is obtained. The training performance score of the LSTM model is obtained for each value within the specified range. Based on the training performance score of the LSTM model for each value within the specified range, the optimal number of hidden units is obtained. The optimal number of hidden units is used as the number of hidden units in the LSTM model to obtain an LSTM prediction model for predicting future multidimensional environmental data, which is used for real-time monitoring of the construction environment in the power engineering construction area. The process of obtaining the maximum number of LSTM parameters in the LSTM model for the device deployed and running according to the LSTM model includes: The device on which the LSTM model is deployed and run is designated as the target device. The available memory of the target device is obtained, and the available memory of the target device is calculated as a preset multiple to obtain the maximum number of LSTM parameters in the LSTM model. The step of obtaining a complexity score for the multidimensional historical environmental data sequence based on its variation characteristics includes: PCA principal component analysis is performed on the multidimensional historical environmental data sequence to obtain at least one principal component and its corresponding variance contribution rate. A scree plot is constructed based on each principal component and its corresponding variance contribution rate. The horizontal axis of the scree plot is the number of principal components, and the vertical axis is the variance contribution rate. Obtain the inflection point of the scree plot. In the scree plot, mark the abscissa of the inflection point as the number of principal components. Use the ratio of the number of principal components to the number of dimensions as the independent variable of the sigmoid function to obtain the first complexity score of the multidimensional historical environment data sequence. The multidimensional historical environment data sequence is divided into a preset number of one-dimensional historical environment data sequences according to the dimensions of the multidimensional historical environment data. The information entropy of each one-dimensional historical environment data sequence is obtained. The information entropy of each one-dimensional historical environment data sequence is used as the independent variable of the sigmoid function, and the mean value of the sigmoid function is obtained, which is denoted as the second complexity score of the multidimensional historical environment data sequence. The complexity score of the multidimensional historical environment data sequence is obtained based on the average between the first complexity score and the second complexity score.

2. The method for real-time monitoring of the construction environment of power engineering according to claim 1, characterized in that, The step of obtaining the initial interval for the number of hidden units when using the LSTM model to predict future multidimensional environmental data based on the maximum value includes: Using the maximum value, the formula for calculating the number of LSTM parameters is reversed to obtain the maximum number of hidden units. The number of dimensions of the multidimensional historical environmental data is then obtained. The number of dimensions and the maximum number of hidden units are used to form the initial interval for the number of hidden units when using the LSTM model to predict future multidimensional environmental data.

3. The method for real-time monitoring of the construction environment of power engineering according to claim 2, characterized in that, The step of obtaining the standardized interval for the number of hidden units based on the complexity score and the initial interval includes: Obtain the difference between the maximum number of hidden units and the number of dimensions to get the span value of the initial interval, calculate the product between the span value and the complexity score, and obtain the sum of the product and the number of dimensions to get the number of normalized hidden units; The fluctuation range of the specified number of hidden units is set as the specified range for the number of hidden units.

4. The method for real-time monitoring of the construction environment of power engineering according to claim 1, characterized in that, The step of training the LSTM model using the multidimensional historical environment data sequence according to a preset training round, and obtaining the training performance score of the LSTM model for any given value, includes: According to a preset ratio, the multidimensional historical environment data sequence is divided into a multidimensional training data sequence and a multidimensional verification data sequence; For any training epoch, the multidimensional training data sequence is input into the LSTM model for training to obtain the adjusted LSTM model for that training epoch. The average error generated by the LSTM model during training is obtained and recorded as the training error. The multidimensional validation data sequence is input into the adjusted LSTM model for validation to obtain the average error generated by the adjusted LSTM model during validation and recorded as the validation error. The mean between the training error and the validation error is calculated to obtain the overall error value for that training epoch. Obtain the training error, validation error, and total error value for each training round. Based on the training error, validation error, and total error value for each training round, obtain the training performance score of the LSTM model for any given value.

5. The method for real-time monitoring of the construction environment of power engineering according to claim 4, characterized in that, The step of obtaining the training performance score of the LSTM model for any given value based on the training error, validation error, and overall error value of each training epoch includes: The total error values ​​of each training round are combined into a total error value sequence. The total error value sequence is fitted to obtain a fitted curve. The inflection point of the fitted curve is obtained and recorded as the total error inflection point. The training round corresponding to the total error inflection point is recorded as the training convergence round. Substitute the total error value of the last training round into the exponential function with the training convergence round as the base to obtain the exponential function result. Use the exponential function result as the independent variable of the sigmoid function to obtain the corresponding sigmoid function value. Obtain the difference between the constant 1 and the sigmoid function value to obtain the first training effect score. Obtain the absolute value of the difference between the training error and the validation error in each training round, and obtain the cumulative value of the absolute value of the difference. Use the cumulative value of the absolute value of the difference as the independent variable of the sigmoid function to obtain the total difference result. Obtain the difference between the constant 1 and the total difference result to obtain the second training effect score. The training performance score of the LSTM model for any given value is obtained by averaging the first training performance score and the second training performance score.

6. The method for real-time monitoring of the construction environment of power engineering according to claim 1, characterized in that, The step of obtaining the optimal number of hidden units based on the training performance score of the LSTM model for each value within the specified range includes: The optimal number of hidden units is determined by taking the number of hidden units corresponding to the maximum training performance score within the specified range.

Images