Power grid load prediction method based on multi-strategy adaptive blatta orientalis optimization deep network and related device

Abstract

The application discloses a power grid load prediction method based on a multi-strategy adaptive melolontha optimization deep network and related devices. In view of the problems that the existing deep prediction network hyperparameter setting depends on artificial experience and the conventional algorithm is prone to local optimal deadlock, the application acquires historical power grid load data to construct a deep prediction network and determine a hyperparameter search boundary; an initial melolontha population is generated by using a Logistic-Tent chaotic mapping to eliminate a space blind area; a root mean square error of a verification set is used as a fitness objective function to drive iteration; in the iteration, a nonlinear adaptive inertia weight and a golden sine function are combined to reconstruct a position update equation, so that the balance between wide area exploration and microscopic convergence is achieved; when optimization stagnation occurs, a t-distribution disturbance and an adaptive Gaussian-Cauchy mixed joint mutation mechanism are triggered to tear the local optimal barrier; finally, a global optimal hyperparameter matrix is output to the deep prediction network to perform power grid load prediction. The application effectively breaks the multi-peak function gradient deadlock and significantly improves the accuracy and robustness of power grid time series load prediction.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and smart grid, and relates to a power grid load forecasting method and related device based on multi-strategy adaptive dung beetle optimization deep network. Background Technology

[0002] In the process of widely applying deep learning and artificial intelligence technologies to industrial time-series data analysis, equipment fault diagnosis, and power grid load forecasting, constructing complex hybrid networks with excellent theoretical topologies (such as deep combined architectures like CNN-BiLSTM) is merely the foundational stage of model development. In fact, these deep neural networks contain an extremely large parameter search space. Besides the tens of millions of weight biases that the network learns spontaneously during backpropagation, there are numerous hyperparameters that determine the network's micro-computational structure and learning state. Even small perturbations to these hyperparameters can cause drastic deformation of the network's high-dimensional non-convex loss surface, directly leading to model deadlock in local minima, uncontrolled convergence time, or a collapse in generalization ability.

[0003] In the current field of hyperparameter optimization for deep networks, existing technologies exhibit significant performance bottlenecks and algorithmic flaws, making it difficult to meet the high-precision tuning requirements at the engineering level. The most traditional hyperparameter tuning methods rely on manual trial and error or exhaustive grid search. When facing multidimensional continuous hyperparameter spaces, grid search is prone to the "curse of dimensionality" due to the complex nonlinear coupling between different parameters, consuming massive amounts of computing power and time, often only reaching suboptimal solutions. The data characteristics of time-series systems such as power grids dynamically drift with seasons and economic cycles, often rendering statically set manual experience parameters completely ineffective in practical engineering applications. To overcome the computational bottleneck of manual tuning, swarm intelligence metaheuristic algorithms (such as Particle Swarm Optimization (PSO), Sine-Cosine Algorithm (SCA), and Bat Algorithm (BA)) have been widely introduced in recent years to automatically take over the task of network hyperparameter tuning. However, these classic heuristic algorithms all suffer from premature convergence and late-stage optimization stagnation when approximating the loss function of highly non-convex deep networks. Even the newly proposed and highly regarded standard Dung Beetle Optimization (DBO) algorithm still suffers from the following three significant structural flaws when dealing with the hyperparameter mapping space of deep networks: Blind spot defects in the initialization stage: The standard algorithm relies too much on the pseudo-random number generator in the population initialization stage, which leads to extremely uneven distribution of dung beetle individuals in the multidimensional solution space, causing serious local clustering effects and generating huge search blind spots, resulting in the loss of the population's inherent diversity.

[0004] Imbalance between breadth and depth of exploration: In the iterative process, the standard algorithm lacks a nonlinear damping mechanism to balance global breadth exploration and local depth mining, which makes it difficult for the search step size to converge effectively in the later stages of iteration, resulting in the optimization accuracy failing to meet the requirements.

[0005] Deadlock in multimodal functions: When an algorithm encounters multimodal terrain, it often loses its search momentum at pseudo-extreme points. Due to the lack of a strong mutation mechanism to break through local optimum traps, the algorithm can only oscillate around the smooth gradient in place, completely losing the ability to jump out of suboptimal solutions and achieve self-deconstruction and reconstruction.

[0006] In summary, overcoming the technical shortcomings of standard swarm intelligent optimization models, such as large population distribution blind spots and easy entrapment in high-dimensional parameter local optima, and breaking through the technical bottleneck of existing algorithms' inability to adaptively balance "large-span escape" and "high-precision local convergence" in complex non-convex optimization spaces, and introducing a novel optimization mechanism with self-correction and multiple escape strategies to automatically and accurately tune the hyperparameters of deep neural networks and successfully apply it to power grid load forecasting systems, has become an urgent technical challenge in this field. Summary of the Invention

[0007] To address the shortcomings of existing deep neural networks in time-series load forecasting, such as the time-consuming manual trial-and-error of hyperparameters, the susceptibility of conventional swarm intelligence optimization algorithms to local minima and deadlock in high-dimensional non-convex parameter spaces, large population distribution blind spots, and late-stage optimization stagnation, this invention aims to provide a power grid load forecasting method and related apparatus based on a multi-strategy adaptive dung beetle optimization deep network. This method deeply coordinates chaos theory, nonlinear kinetic energy regulation, and a dual mutation mechanism over time to construct a parameter optimization closed loop with self-correction and multiple escape capabilities, objectively and efficiently improving the accuracy and robustness of short-term power grid load forecasting.

[0008] This invention is achieved through the following technical solution: A power grid load forecasting method based on a multi-strategy adaptive dung beetle optimized deep network includes, Historical power grid operating load data is acquired and divided into training and validation sets. A deep prediction network is constructed and the search boundary of the hyperparameters to be optimized is determined. The initial dung beetle population was generated within the hyperparameter search boundary using the Logistic-Tent chaotic mapping, thus constructing the initial hyperparameter solution space; The root mean square error of the deep prediction network on the validation set is used as the objective function of environmental fitness to drive the population to perform optimization iteration. During the population iterative update process, the position update equations of each individual in the first generation dung beetle population are reconstructed by combining an adaptive inertial weight that decays nonlinearly with the number of iterations and a sine function containing a constant golden ratio. When the deviation of the environmental fitness objective function continuously approaches zero, a t-distribution perturbation based on the current iteration number is applied to the individuals in the population to determine the current global optimal individual. Then, a Gaussian-Cauchy hybrid joint mutation reconstruction based on adaptive probability weights is applied. After reaching the maximum number of iterations, the global optimal hyperparameter matrix is ​​output. The globally optimal hyperparameter matrix is ​​configured into the deep prediction network, and the network is trained using the training set. Then, the current power grid data is input, and the predicted power grid load value is output.

[0009] Preferably, the hyperparameters to be optimized include: number of convolutional kernels, kernel size, number of bidirectional hidden layer nodes, initial learning rate, and batch size.

[0010] Preferably, in the step of calculating and generating the initial dung beetle population using the Logistic-Tent chaotic mapping within the hyperparameter search boundary, the one-dimensional chaotic update equation of the Logistic-Tent chaotic mapping is:

[0011] In the formula, Here, r is the nth generation chaotic time series value, and r is a control parameter with a value of 3.99. After calculating and generating the chaotic time series value, the chaotic time series value is back-mapped to the set hyperparameter search boundary, wherein the boundary range of the initial learning rate is limited to [0.0001, 0.001].

[0012] Preferably, in the step of reconstructing the position update equations for each individual in the initial dung beetle population, the adaptive inertia weight is calculated using the following formula: ; In the formula, Let t be the adaptive inertia weight, and t be the current iteration number. For the maximum number of iterations, The maximum weight value is set. This is the minimum weight value set. The position update equations for each individual in the initial dung beetle population are reconstructed as follows:

[0013] In the formula, Let i be the updated position of the i-th individual. Let i be the current position of the i-th individual. It is [0, The random number within [ ] controls the step size direction. The current position of the globally optimal individual, and the space reduction coefficient. ,in It is a constant of the golden ratio and satisfies .

[0014] Preferably, a t-distribution perturbation based on the current iteration number is applied to each of the individuals, and the formula for calculating the t-distribution perturbation is:

[0015] In the formula, The new position after the disturbance. The current position of each of the entities. It is a t-distributed random mutation operator with the current iteration number as the degree of freedom index.

[0016] Preferably, in the step of performing mutation reconstruction and position update on the globally optimal individual, the mutation reconstruction formula is:

[0017] In the formula, This represents the new location region of the globally optimal individual after the mutation reconstruction. This represents the current position of the globally optimal individual. The standard Cauchy distribution function is used. It is the standard Gaussian distribution function; and The adaptive probability weights are dynamically changed with each iteration round, and satisfy the following conditions: Where t is a brief description of the current iteration number, The maximum number of iterations.

[0018] Preferably, the initial dung beetle population includes rolling dung beetles, reproductive dung beetles, foraging dung beetles, and stealing dung beetles. Their respective basic displacement equations include rolling behavior influenced by the global worst-case position, reproductive behavior seeking optimization within the safe boundary, foraging behavior performing local precision mining, and stealing behavior approaching the global optimal solution. Specifically: The displacement equation of the dung beetle is: ,in ,in, This is the worst position globally. Let i be the updated position of the i-th individual. Let i be the current position of the i-th individual. This refers to the positional deviation affected by the worst-case global position. , , These are the basic displacement bias coefficients and natural weight constants representing the behavior of being pulled by the light source and avoiding obstacles in the standard dung beetle algorithm; The displacement equation of the reproductive dung beetle is: ,in It is a local extremum point; To find the optimal lower bound for reproductive behavior within the safe boundary. The upper bound for finding the optimal safety boundary for reproductive behavior within the safety boundary. A constant factor used to control the position update of dung beetles towards local extrema and safe boundaries; The displacement equation of the foraging dung beetle is: ;in, These are the lower and upper boundaries of the foraging area where the dung beetle performs localized, precise mining, respectively. To control the coefficients for foraging dung beetles to update their location and dig at different depths within a local area; The displacement equation of the dung beetle is: ;in, The position of the globally optimal individual; The respective actions are to control the dung beetle to quickly move towards the current global optimal solution. and local extreme points The weighting step size coefficient for convergence.

[0019] A power grid load forecasting device based on a multi-strategy adaptive dung beetle optimized deep network includes: The data and network construction module is used to acquire historical power grid operating load data and divide it into training and validation sets, construct a deep prediction network, and determine the search boundary of the hyperparameters to be optimized. The chaos initialization module is used to generate the first generation of dung beetle population within the hyperparameter search boundary using the Logistic-Tent chaotic mapping, and to construct the initial hyperparameter solution space; The fitness optimization module is used to use the root mean square error of the deep prediction network on the validation set as the environmental fitness objective function to drive the first generation dung beetle population to perform optimization iteration; The position equation reconstruction module is used to reconstruct the position update equation of each individual in the first generation of the mantis population by combining an adaptive inertial weight that decays nonlinearly with the number of iterations and a sine function containing a constant golden ratio during the population iteration update process. The mutation escape and output module is used to apply a t-distribution perturbation based on the current iteration number to the individuals in the population when the gap of the environmental fitness objective function continuously approaches zero, to determine the current global optimal individual, and to apply Gaussian-Cauchy hybrid joint mutation reconstruction based on adaptive probability weights. After reaching the maximum number of iterations, the module outputs the global optimal hyperparameter matrix. The forward load forecasting module is used to configure the globally optimal hyperparameter matrix into the deep forecasting network, complete network training using the training set, and then input the current power grid data to output the power grid load forecast value.

[0020] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a power grid load forecasting method based on a multi-strategy adaptive dung beetle optimized deep network.

[0021] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the described method for power grid load forecasting based on a multi-strategy adaptive dung beetle optimized deep network.

[0022] Compared with the prior art, the present invention has the following beneficial technical effects: To address the shortcomings of existing deep neural networks in time-series load forecasting, such as the time-consuming manual trial and error of hyperparameters, the susceptibility of conventional swarm intelligence optimization algorithms to local minima and deadlocks in high-dimensional non-convex parameter spaces, large population distribution blind spots, and late-stage optimization stagnation, this invention aims to provide a power grid load forecasting method and related devices based on a multi-strategy adaptive dung beetle optimization deep network. This method deeply coordinates chaos theory, nonlinear kinetic energy regulation, and a dual mutation mechanism over time to construct a parameter optimization closed loop with self-correction and multiple escape capabilities, objectively and efficiently improving the accuracy and robustness of short-term power grid load forecasting. This invention constructs a multi-strategy adaptive dung beetle optimization algorithm (MSADBO) by functionally coupling and coordinating multiple strategies on the time axis, including Logistic-Tent chaotic initialization, nonlinear adaptive weights and golden sine iteration, t-distribution perturbation, and Gaussian-Cauchy hybrid mutation. This scheme achieves a synergistic effect of "1+1>2" among the distinguishing features, overcomes the inherent structural defects of standard swarm intelligence algorithms in complex non-convex optimization spaces, and objectively and accurately improves the efficiency and prediction accuracy of hyperparameter tuning in deep neural networks. Furthermore, this invention abandons the conventional pseudo-random initialization method and utilizes the ergodicity of the Logistic-Tent hybrid chaotic mapping to generate chaotic time series values ​​and map them to the hyperparameter boundary. This ensures, from the physical source of the population distribution mechanism, that the first-generation dung beetle individuals can be uniformly distributed in the high-dimensional non-convex solution space, effectively avoiding local clustering and trap convergence problems caused by the loss of inherent population diversity, and significantly expanding the initial optimization scope of the algorithm.

[0023] Furthermore, this invention constructs an adaptive inertia weight that decays nonlinearly with the number of iterations and nests it into a sinusoidal periodic oscillation operator incorporating a constant golden ratio for position update equation reconstruction. This collaborative mechanism enables the algorithm to utilize large sinusoidal amplitudes to provide high kinetic energy in the early stages of iteration, quickly traversing large step sizes to lock onto extreme value basins; in the later stages of iteration, it achieves high-precision parameter fine-tuning through nonlinear damping and steep descent, relying on extremely low inertia. This solves the technical challenge of a single constant step size failing to simultaneously address global exploration and local depth mining, demonstrating extremely accurate convergence guidance capabilities in challenging terrains (such as the Rosenbrock benchmark). This invention not only provides the underlying algorithmic theory but also further encapsulates it with deep prediction networks into a complete business system using a B / S architecture. By providing an isolated dual-role authentication system, persistent data storage, and a one-click rollback mechanism for model configuration, it transforms complex hyperparameter tuning logic into automated microservices, eliminating the technical barriers for power dispatchers operating the underlying intelligent code and realizing a direct transformation from algorithm optimization to actual industrial dispatch productivity.

[0024] Furthermore, this invention addresses the stagnation problem of optimization algorithms easily losing search momentum in the later stages of approximating highly non-convex loss functions. It introduces a dual escape mechanism based on the dynamic evolution of the t-distribution during iterative degrees of freedom and a combined Gaussian-Cauchy mutation. When the fitness function drop approaches zero, a large gradient jump is generated using the peak and extremely long tail of the Cauchy distribution at the zero point, instantly breaking through the local optimum barrier of multimodal functions (such as the Ackley function). Simultaneously, the autonomous evolution of the t-distribution with each iteration achieves a smooth transition, ensuring that the population can quickly and stably converge after escaping the local extremum trap.

[0025] Furthermore, this invention configures the automatically optimized globally optimal hyperparameter matrix output by the power grid short-term load composite prediction network (such as VMD-CNN-Attention-BiLSTM) for forward inference prediction. Without altering any physical connections at the grid neuron level, driven by the optimal logical topology, the root mean square error (RMSE) of the validation set is substantially reduced from the unoptimized empirical value of 117.86 to 105.63, and the coefficient of determination (R²) jumps to 0.9504. This fully demonstrates the high accuracy and robustness of this multi-strategy optimization mechanism compared to existing baseline models in actual power grid load tracking. Attached Figure Description To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of the power grid load forecasting method based on a multi-strategy adaptive dung beetle optimized deep network according to the present invention. Figure 2 The graph shows the convergence curves of each optimization algorithm in Example 1 on the test function. Figure 3 This is a comparison chart of load forecasting results after adding the optimization algorithm to Example 1; Figure 4 The above describes the overall architecture of the power prediction system in Example 1. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0028] A power grid load forecasting method based on a multi-strategy adaptive dung beetle optimization deep network, such as... Figure 1 As shown, including, S1. Obtain historical power grid operating load data and divide it into training set and validation set, construct a deep prediction network and determine the search boundary of hyperparameters to be optimized; The specific process is as follows: Collect the actual historical operating load time series data of a certain city's power grid, perform necessary data preprocessing, and then divide it into a training set for model fitting and a validation set for calculating the root mean square error (RMSE). The actual historical operating load time series data of the power grid specifically includes multi-dimensional feature information that dynamically drifts with the seasons and economic cycles. Its core not only covers the actual daily electricity load value at the corresponding historical sampling time, but also deeply integrates the exogenous variable features that have a significant impact on load fluctuations, including meteorological environmental feature data such as temperature, humidity, and wind speed, as well as time calendar labels including specific sampling times, distinctions between weekdays and weekends, and identification of statutory holidays, thus jointly constituting a multi-dimensional input feature set that can comprehensively characterize the complex dynamic characteristics of the power grid time series system.

[0029] The specific process of constructing a deep prediction network is as follows: On a high-performance computing node, a deep neural network is independently instantiated as an optimized business carrier. A complex hybrid network base with a good theoretical topology is defined and constructed. The network structure is a deep composite architecture composed of modules such as one-dimensional convolutional neural network (1D-CNN) and bidirectional long short-term memory network (BiLSTM) (such as VMD-CNN-Attention-BiLSTM model). Without changing the physical connection of the grid neurons, the network is allowed to spontaneously learn tens of millions of weight biases during the backpropagation process of training. Finally, relying on the logical topology set at the bottom of the architecture, the network receives the grid time series data and completes the deep extraction of complex time series features and forward inference prediction of the short-term daily load of the grid.

[0030] For the aforementioned deep prediction network, the 5-dimensional continuous hyperparameter search boundary that determines its micro-computational structure and learning state is objectively delineated. Specifically, the parameter domain to be optimized is defined as the number of convolutional kernels, kernel size, number of bidirectional hidden layer nodes, initial learning rate, and batch size, thereby constructing an initial optimization solution space with clear physical meaning.

[0031] S2. The initial dung beetle population is generated within the hyperparameter search boundary using Logistic-Tent chaotic mapping, constructing the initial hyperparameter solution space. Specifically, to overcome the shortcomings of traditional optimizers using pseudo-random number generators, which result in extremely uneven distribution and local clustering effects in the multidimensional solution space, a one-dimensional hybrid chaotic update equation based on Logistic-Tent is introduced. The control parameter r of this equation is assigned a value of 3.99 to calculate and generate chaotic time-series values ​​with high ergodicity and non-repetition. Subsequently, the generated chaotic time-series values ​​are back-mapped to the pre-defined physical boundary of the deep network hyperparameters (e.g., the initial learning rate is limited to [0.0001, 0.001]), thereby generating several (e.g., 30) initial dung beetle individuals carrying different parameter combinations and extremely uniformly distributed within the non-convex solution space. This maximizes the inherent diversity of the initial parameter combinations and completely eliminates the high-dimensional search blind zone, completing the construction of a high-quality initial hyperparameter solution space.

[0032] S3, The root mean square error of the deep prediction network on the validation set is used as the objective function of environmental fitness to drive the population to perform optimization iteration; S4. During the population iterative update process, the position update equations of each individual in the first generation dung beetle population are reconstructed by combining the adaptive inertial weight that decays nonlinearly with the number of iterations and the sine function containing the constant of the golden ratio. The specific process is as follows: To avoid premature convergence, the constant iteration step size is first abandoned, and an adaptive inertial weight that decays nonlinearly with the number of iterations is constructed as a dynamic valve. Subsequently, this decaying kinetic energy is innovatively injected into and nested into a periodic oscillation operator that incorporates a sine function with a constant golden ratio, thus completely reconstructing the individual position update paradigm responsible for wide-area exploration and local oviposition operations in the standard dung beetle optimization algorithm. In the reconstructed position update equation, the step size is controlled by random numbers within the interval, and combined with the spatial reduction coefficient calculated by the golden ratio, which is equivalent to equipping the prediction model with an intelligent continuously variable transmission. This allows the entire hyperparameter optimization process to maintain strong kinetic energy in the early stage by relying on large sinusoidal amplitude, thereby quickly locking the extreme value basin and crossing the flat non-optimal region by traversing large step sizes. In the later stage of iteration, the kinetic energy drops sharply, and the convergence constant and extremely low inertia are used to firmly hold onto the optimal neighborhood to perform microsecond-level high-precision parameter fine-tuning and in-depth exploration.

[0033] S5, When the gap of the environmental fitness objective function continuously approaches zero, apply a t-distribution perturbation based on the current iteration number to the individuals in the population, determine the current global optimal individual, and apply Gaussian-Cauchy hybrid joint mutation reconstruction based on adaptive probability weights. After reaching the maximum number of iterations, output the global optimal hyperparameter matrix. The specific process is as follows: when the difference in the environmental fitness objective function over multiple generations approaches zero and does not show a significant decrease, it is determined that the optimization system has fallen into the local trap of the multi-peak function and gradient deadlock. At this time, the dual escape escape operator is forcibly activated. Using the current number of algorithm iterations as the core degree of freedom variable, a t-distribution perturbation is applied to the individuals in the population, so that their mutation characteristics are dynamically switched as the optimization process progresses. That is, in the early stage of iteration, the long-tail characteristics of the Cauchy distribution are simulated to provide strong mutation leaps, while in the later stage, the characteristics of the Gaussian distribution are approximated to guide the fine adjustment and smooth transition of the local domain. The algorithm redefines and performs an extreme hybrid joint mutation reconstruction on the stuck current global optimal leader. It uses adaptive probability weights that dynamically change with each iteration to fuse the perturbation vectors of the standard Gaussian distribution and the Cauchy distribution, which are then directly applied to the optimal hyperparameter solution. Relying on the peak and extremely long tail of the Cauchy distribution at the zero point, it generates a large gradient jump, which powerfully and instantaneously tears apart the local optimal barrier generated by the complex topology of the deep network. After undergoing the above mutation mechanism and reaching the set maximum number of iterations, the algorithm completely converges and terminates the iteration, finally outputting the optimal deep network hyperparameter matrix corresponding to the global minimization error.

[0034] S6. Configure the globally optimal hyperparameter matrix into the deep prediction network, and complete the network training using the training set. Then, input the current power grid data and output the power grid load prediction value.

[0035] The specific process is as follows: The automatically generated globally optimal hyperparameter matrix, which has been stably output after optimization through multiple strategies, is directly injected and configured into the aforementioned composite deep prediction network base. Without altering any physical connections at the grid neuron level, the optimal logical topology and micro-computing power structure of the model are established. Subsequently, the preprocessed historical load training set of the power grid is input into the configured network. Through the underlying backpropagation mechanism, the network learns spontaneously and continuously updates the tens of millions of weight biases within the network until the model loss function fully converges, thus completing the final instantiation training of the deep prediction network. When a prediction and scheduling task is generated in the actual power grid operation, the current real-time feature data stream of the power grid, containing multi-dimensional information such as the latest meteorological characteristics and time labels, is input into the trained and solidified prediction model. The system executes the forward inference process with one click, calculates and outputs the predicted power grid load value for the future period, and can use the front-end visualization chart library to compare and render the predicted future load line with the historical real curve and export standard scheduling reports, thereby completely realizing the direct transformation from the underlying deep learning optimization algorithm to the actual productivity of the power grid.

[0036] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0037] Example 1: A Power Grid Load Forecasting Method Based on Multi-Strategy Adaptive Dung Beetle Optimization Deep Network To address the technical challenges faced by existing deep neural networks in power grid load forecasting, such as time-consuming manual trial-and-error methods and susceptibility to deadlock in local minima in high-dimensional non-convex parameter spaces, this embodiment provides a high-precision automated optimization and forecasting method. The specific implementation steps are as follows: Step S1: Data Processing and Prediction Foundation Construction Historical power grid operating load data is obtained, and after cleaning and normalization preprocessing, the data is divided into training set and validation set.

[0038] Construct a deep prediction network for feature extraction and temporal prediction (preferably a composite architecture including a one-dimensional convolutional neural network 1D-CNN and a bidirectional long short-term memory network BiLSTM in this embodiment).

[0039] The search boundary for the 5-dimensional continuous hyperparameters to be optimized in this deep prediction network is determined, and the hyperparameters include: the number of convolutional kernels. Core size Bidirectional hidden layer node Initial learning rate and batch size ; Step S2: Construct an initial population spatial distribution mechanism based on the Logistic-Tent hybrid chaotic mapping.

[0040] This invention abandons the pseudo-random number generator used in traditional optimizers and utilizes the ergodicity and non-repetition properties of chaos theory to generate uniformly distributed initial dung beetle individuals within a defined boundary space of deep network hyperparameters. This maximizes the diversity of initial parameter combinations and completely eliminates blind spots in high-dimensional search. This invention abandons conventional random initialization and employs the ergodicity and non-repetition properties of chaos theory for population distribution.

[0041] Chaotic time series values ​​are generated using the Logistic-Tent one-dimensional chaotic update equation. The specific formula is as follows:

[0042] In the formula, Here, r is the nth generation chaotic time series value, and r is a control parameter with a value of 3.99. After calculating and generating the chaotic time series value, the chaotic time series value is back-mapped to the set hyperparameter search boundary, wherein the boundary range of the initial learning rate is limited to [0.0001, 0.001].

[0043] The generated chaotic time series values ​​are back-mapped to the hyperparameter physical boundary set in step S1, generating several (e.g., 30) initial dung beetle individuals carrying different parameter combinations, thus constructing the initial hyperparameter solution space. Taking the initial learning rate as an example, it is mapped and limited to the physical boundary range of [0.0001, 0.001]. This mechanism significantly expands the initial optimization field of view, eliminating local trap convergence caused by blind spots from the physical source.

[0044] Step S3: Construct a deep learning prediction network base and define the fitness evaluation function Define the deep network architecture to be optimized (e.g., a composite of 1D-CNN and BiLSTM), and use the root mean square error (RMSE) of the deep prediction network on the validation set as the environment fitness objective function. Run the network using individuals with initialized hyperparameters, calculate the prediction residuals for each set of parameter configurations, and determine the current globally optimal hyperparameter position. Location of the global worst hyperparameter .

[0045] Within the population, basic role assignments and position updates are performed according to the mechanism of the standard dung beetle optimization algorithm, including: The breeding dung beetle, simulated as being pulled by a light source and avoiding obstacles, has the following displacement equation: ,in It is a local extremum point; To find the optimal lower bound for reproductive behavior within the safe boundary. The upper bound for finding the optimal safety boundary for reproductive behavior within the safety boundary. A constant factor used to control the position update of dung beetles towards local extrema and safe boundaries; The foraging dung beetle seeks optimally within the safety boundary, and its displacement equation is: ;in, These are the lower and upper boundaries of the foraging area where the dung beetle performs localized, precise mining, respectively. To control the coefficients for foraging dung beetles to update their location and dig at different depths within a local area; The dung beetle that steals minerals performs localized precision mining; its displacement equation is: ;in, The position of the globally optimal individual; The respective actions are to control the dung beetle to quickly move towards the current global optimal solution. and local extreme points The weighting step size coefficient for convergence.

[0046] The dung beetle, as it approaches the globally optimal solution, has the following displacement equation: ,in It is a local extremum point; To find the optimal lower bound for reproductive behavior within the safe boundary. The upper bound for finding the optimal safety boundary for reproductive behavior within the safety boundary. A constant factor used to control the position update of dung beetles towards local extrema and safe boundaries; The displacement equation of the foraging dung beetle is: ;in, These are the lower and upper boundaries of the foraging area where the dung beetle performs localized, precise mining, respectively. To control the coefficients for foraging dung beetles to update their location and dig at different depths within a local area; The displacement equation of the dung beetle is: ;in, The position of the globally optimal individual; The respective actions are to control the dung beetle to quickly move towards the current global optimal solution. and local extreme points The weighting step size coefficient for convergence.

[0047] Step S4: Extensive and In-Depth Mining of Nonlinear Inertia and the Golden Sine Wave A nonlinear adaptive inertial weight and a golden sine iteration mechanism are introduced to drive the population to perform breadth-of-field search and depth-of-field mining. An adaptive inertial weight with nonlinear smooth decay over iterations controls the exploration kinetic energy. Simultaneously, the golden ratio constant is embedded into the periodic oscillation operator of the sine function, reconstructing the position update equation responsible for wide-area exploration and local spawning operations in the standard algorithm. This forces the entire parameter optimization process to quickly lock onto extreme value basins with large step sizes in the early stages, and achieves extremely high-precision parameter fine-tuning in the later stages through nonlinear decay.

[0048] To avoid the premature convergence problem of conventional heuristic algorithms, this invention constructs an adaptive inertia weight that decays nonlinearly and smoothly with the current iteration number t. : ; In the formula, Let t be the adaptive inertia weight, and t be the current iteration number. For the maximum number of iterations, The maximum weight value is set. This is the minimum weight value set. The decaying kinetic energy is incorporated into a periodic oscillation operator of a sinusoidal function containing a constant golden ratio, and the position update equation is reconstructed for the individuals responsible for wide-area expansion and local spawning operations in step S3. The reconstructed equation is:

[0049] In the formula, Let i be the updated position of the i-th individual. Let i be the current position of the i-th individual. It is [0, The random number within [ ] controls the step size direction. The current position of the globally optimal individual, and the space reduction coefficient. ,in It is a constant of the golden ratio and satisfies This mechanism enables the entire parameter optimization process to quickly lock in the extreme value basin in the early stage by relying on the large amplitude of sinusoidal waves and the large step size, and to achieve extremely high precision parameter fine-tuning in the later stage by relying on the extremely low inertia to dig deep and accurate.

[0050] Step S5: Gaussian-Cauchy Destruction Reconstruction Based on Dual Escape Mechanism A perturbation escape mechanism based on a dynamically evolving t-distribution (Student's t-distribution) is constructed. Using the current number of algorithm iterations as the core degree of freedom variable, the mutation distribution characteristics are dynamically switched as the optimization process progresses. In the early stages of iteration, the long-tail characteristics of the Cauchy distribution are simulated, providing strong mutation leaps; in the later stages, the characteristics of the Gaussian distribution are approximated, guiding fine-tuning in local domains.

[0051] A Gaussian-Cauchy hybrid joint mutation reconstruction is implemented for the globally optimal solution. When the fitness function fails to decrease significantly over multiple generations, indicating the system is trapped in a local trap of a multimodal function, extreme intervention is triggered. Adaptive probability weights are applied to fuse the perturbation vectors of Gaussian and Cauchy distributions, directly acting on the current optimal hyperparameter solution to break local gradient deadlock using long-tail mutations. After the maximum number of iterations, the optimal deep network hyperparameter matrix corresponding to the global minimization error is output.

[0052] When the system monitors that the difference in the environmental fitness objective function (RMSE) over multiple generations approaches zero without significant decrease, it determines that the system is trapped in a local gradient deadlock of a multimodal function and forcibly activates the dual escape operator.

[0053] Using the current algorithm iteration number t as the core degree of freedom variable, apply Student's t-distribution perturbation to the individual:

[0054] In the formula, The new position after the disturbance. The current position of each of the entities. The t-distribution random mutation operator, with the current iteration number as the degree of freedom index, simulates the long-tail characteristics of the Cauchy distribution in the early stage of iteration to provide strong mutation, and in the later stage, it approximates the Gaussian distribution to guide local fine adjustments.

[0055] For the stuck global optimal leader individual Perform Gauss-Cauchy joint destruction and reconstruction:

[0056] In the formula, This represents the new location region of the globally optimal individual after the mutation reconstruction. This represents the current position of the globally optimal individual. The standard Cauchy distribution function is used. It is the standard Gaussian distribution function; and The adaptive probability weights are dynamically changed with each iteration round, and satisfy the following conditions: Where t is a brief description of the current iteration number, The maximum number of iterations is given. Utilizing the mathematical property that the Cauchy part has a sharp peak at zero and extremely long tails at both ends, a large gradient jump is generated, instantly breaking through the local optimum barrier (such as the trap of the Ackley function). Step S6: Parameter Output and Business Application Mapping After the maximum number of iterations, the algorithm converges and outputs the optimal hyperparameter matrix of the deep network corresponding to the global minimum error. This optimal matrix is ​​then configured into the deep prediction network for model solidification, and the final training is completed using the training set data. Experimental data shows that this optimal logical topology significantly reduces the root mean square error (RMSE) of the network from the empirical value of 117.86 to 105.63, and the coefficient of determination (R²) also decreases. 2 The value jumped to 0.9504. When a power grid dispatching task is generated, real-time power grid characteristic data is input, and forward inference is used to output power grid load forecast lines and reports for future periods, realizing the direct transformation of the underlying deep learning algorithm into actual productivity.

[0057] This invention constructs a business system using a front-end and back-end separation framework, persistently stores the optimal model version in a relational database, and transforms the complex optimization logic into a scheduling tool through one-click forward inference and visualization chart rendering. This not only eliminates the code barrier for power dispatchers to operate the underlying algorithms but also realizes the direct transformation from intelligent optimization theory to actual power grid productivity.

[0058] like Figure 2 As shown in the convergence curves of the four benchmark test functions, the improved Multi-Strategy Adaptive Dung Beetle Optimization (MSADBO) algorithm achieves a significant leap forward in both optimization accuracy and the ability to escape local optima compared to the original DBO algorithm. Although its absolute convergence speed in simple unimodal functions (F1) is slower than PSO, MSADBO demonstrates strong robustness and excellent global exploration potential when facing complex multimodal terrains (F2, F3, F4) that are prone to local deadlock. In particular, the frequent "step-like" large jumps triggered in the later stages of iteration fully demonstrate that the multi-strategy fusion mechanism can effectively maintain population diversity and successfully break the local optimum trap, resulting in its high-precision optimization performance being superior to the original algorithm.

[0059] like Figure 3As shown in the comparison curves of load forecasting results with and without the optimization algorithm, the MSADBO-optimized forecasting model (red solid line) shows the highest fit to the actual daily electricity load (black solid line) throughout the entire time series, compared to the unoptimized basic VMD-CNN-Attention-BiLSTM model (green dashed line) and the model optimized with standard DBO (blue dotted line). Especially during peaks, troughs, and periods of significant load fluctuation, the MSADBO-optimized red line more accurately and closely tracks the actual load change trend, significantly reducing prediction bias at extreme points. This intuitive comparison fully demonstrates that the multi-strategy adaptive dung beetle optimization algorithm can more efficiently and stably optimize the hyperparameters of this hybrid deep learning architecture globally, thereby effectively improving the overall accuracy and robustness of short-term daily load forecasting for the power grid.

[0060] Figure 4 The overall architecture of the invented power forecasting system is demonstrated, employing a top-down, four-layer modular design. The top layer is the role layer, clearly defining the user permission boundaries between ordinary roles and administrators. Below this is the data layer, responsible for providing comprehensive data lifecycle management functions such as data uploading, viewing, deletion, and preprocessing. The core forecasting task layer constructs a flexible modeling workflow, allowing users to customize the forecasting dataset, forecasting model, optimization algorithm, and model hyperparameters. The bottom result analysis layer is used for in-depth evaluation of the forecasting output, including forecasting result analysis, evaluation index analysis, and forecasting report generation. The overall architecture is logically clear, presenting a complete closed-loop business process from user permission control, data preparation, core model construction to final result output and evaluation.

[0061] This embodiment provides a computer device, including a processor, a memory, and a communication interface.

[0062] The memory stores computer program instructions. When the processor calls the instructions in the memory via the bus, it executes the various steps described in Example 1. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), an application-specific integrated circuit (ASIC), or a field-programmable gate array (FPGA), etc.

[0063] This embodiment also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the operation steps of the building energy consumption online calibration and health monitoring method based on digital twins described in Embodiment 1. The storage medium can be any non-volatile medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0064] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0065] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the accompanying drawings and the above description. However, any modifications, alterations, or variations made by those skilled in the art without departing from the scope of the present invention, utilizing the disclosed technical content, are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A power grid load forecasting method based on a multi-strategy adaptive dung beetle optimization deep network, characterized in that, include, Historical power grid operating load data is acquired and divided into training and validation sets. A deep prediction network is constructed and the search boundary of the hyperparameters to be optimized is determined. The initial dung beetle population was generated within the hyperparameter search boundary using the Logistic-Tent chaotic mapping, thus constructing the initial hyperparameter solution space; The root mean square error of the deep prediction network on the validation set is used as the objective function of environmental fitness to drive the population to perform optimization iteration. During the population iterative update process, the position update equations of each individual in the first generation dung beetle population are reconstructed by combining an adaptive inertial weight that decays nonlinearly with the number of iterations and a sine function containing a constant golden ratio. When the deviation of the environmental fitness objective function continuously approaches zero, a t-distribution perturbation based on the current iteration number is applied to the individuals in the population to determine the current global optimal individual. Then, a Gaussian-Cauchy hybrid joint mutation reconstruction based on adaptive probability weights is applied. After reaching the maximum number of iterations, the global optimal hyperparameter matrix is ​​output. The globally optimal hyperparameter matrix is ​​configured into the deep prediction network, and the network is trained using the training set. Then, the current power grid data is input, and the predicted power grid load value is output.

2. The power grid load forecasting method based on a multi-strategy adaptive dung beetle optimization deep network according to claim 1, characterized in that, The hyperparameters to be optimized include: number of convolutional kernels, kernel size, number of bidirectional hidden layer nodes, initial learning rate, and batch size.

3. The power grid load forecasting method based on a multi-strategy adaptive dung beetle optimization deep network according to claim 1, characterized in that, In the step of calculating and generating the initial dung beetle population using the Logistic-Tent chaotic mapping within the hyperparameter search boundary, the one-dimensional chaotic update equation of the Logistic-Tent chaotic mapping is: In the formula, Here, r is the nth generation chaotic time series value, and r is a control parameter with a value of 3.

99. After calculating and generating the chaotic time series value, the chaotic time series value is back-mapped to the set hyperparameter search boundary, wherein the boundary range of the initial learning rate is limited to [0.0001, 0.001].

4. The power grid load forecasting method based on a multi-strategy adaptive dung beetle optimization deep network according to claim 1, characterized in that, In the step of reconstructing the position update equations for each individual in the initial dung beetle population, the formula for calculating the adaptive inertia weight is as follows: ； In the formula, Let t be the adaptive inertia weight, and t be the current iteration number. For the maximum number of iterations, The maximum weight value is set. This is the minimum weight value set. The position update equations for each individual in the initial dung beetle population are reconstructed as follows: In the formula, Let i be the updated position of the i-th individual. Let i be the current position of the i-th individual. It is [0, The random number within [ ] controls the step size direction. The current position of the globally optimal individual, and the space reduction coefficient. ,in It is a constant of the golden ratio and satisfies .

5. The power grid load forecasting method based on a multi-strategy adaptive dung beetle optimization deep network according to claim 1, characterized in that, A t-distribution perturbation based on the current iteration number is applied to each of the individuals, and the formula for calculating the t-distribution perturbation is: In the formula, The new position after the disturbance. The current position of each of the entities. It is a t-distributed random mutation operator with the current iteration number as the degree of freedom index.

6. The power grid load forecasting method based on a multi-strategy adaptive dung beetle optimization deep network according to claim 1, characterized in that, In the step of performing mutation reconstruction and position update on the globally optimal individual, the mutation reconstruction formula is: In the formula, This represents the new location region of the globally optimal individual after the mutation reconstruction. This represents the current position of the globally optimal individual. The standard Cauchy distribution function, It is the standard Gaussian distribution function; and The adaptive probability weights are dynamically changed with each iteration round, and satisfy the following conditions: Where t is a brief description of the current iteration number, The maximum number of iterations.

7. The power grid load forecasting method based on a multi-strategy adaptive dung beetle optimization deep network according to claim 1, characterized in that, The initial dung beetle population includes rolling dung beetles, reproductive dung beetles, foraging dung beetles, and stealing dung beetles. Their corresponding basic displacement equations include rolling behavior influenced by the global worst-case position, reproductive behavior seeking optimization within the safe boundary, foraging behavior performing local precision mining, and stealing behavior approaching the global optimal solution. Specifically: The displacement equation of the dung beetle is: ,in ,in, This is the worst position globally. Let i be the updated position of the i-th individual. Let i be the current position of the i-th individual. This refers to the positional deviation affected by the worst-case global position. , , These are the basic displacement bias coefficients and natural weight constants representing the behavior of being pulled by the light source and avoiding obstacles in the standard dung beetle algorithm; The displacement equation of the dung beetle is: ,in It is a local extremum point; To find the optimal lower bound for reproductive behavior within the safe boundary. The upper bound for finding the optimal safety boundary for reproductive behavior within the safety boundary. A constant factor used to control the positional updates of dung beetles towards local extrema and safe boundaries; The displacement equation of the foraging dung beetle is: ;in, These are the lower and upper boundaries of the foraging area where the dung beetle performs localized, precise mining, respectively. To control the coefficients for foraging dung beetles to update their location and dig at different depths within a local area; The displacement equation of the dung beetle is: ;in, The position of the globally optimal individual; The respective actions are to control the dung beetle to quickly move towards the current global optimal solution. and local extreme points The weighting step size coefficient for convergence.

8. A power grid load forecasting device based on a multi-strategy adaptive dung beetle optimized deep network, characterized in that, include: The data and network construction module is used to acquire historical power grid operating load data and divide it into training and validation sets, construct a deep prediction network, and determine the search boundary of hyperparameters to be optimized. The chaos initialization module is used to generate the first generation of dung beetle population within the hyperparameter search boundary using the Logistic-Tent chaotic mapping, and to construct the initial hyperparameter solution space; The fitness optimization module is used to use the root mean square error of the deep prediction network on the validation set as the environmental fitness objective function to drive the first generation dung beetle population to perform optimization iteration; The position equation reconstruction module is used to reconstruct the position update equation of each individual in the first generation of the mantis population by combining an adaptive inertial weight that decays nonlinearly with the number of iterations and a sine function containing a constant golden ratio during the population iteration update process. The mutation escape and output module is used to apply a t-distribution perturbation based on the current iteration number to the individuals in the population when the gap of the environmental fitness objective function continuously approaches zero, to determine the current global optimal individual, and to apply Gaussian-Cauchy hybrid joint mutation reconstruction based on adaptive probability weights. After reaching the maximum number of iterations, the module outputs the global optimal hyperparameter matrix. The forward load forecasting module is used to configure the globally optimal hyperparameter matrix into the deep forecasting network, complete network training using the training set, and then input the current power grid data to output the power grid load forecast value.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the power grid load forecasting method based on a multi-strategy adaptive dung beetle optimized deep network as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of a power grid load forecasting method based on a multi-strategy adaptive dung beetle optimized deep network as described in any one of claims 1 to 7.

Images