Method for predicting material demand by deep learning model based on swarm intelligence

Abstract

The invention discloses a method for predicting the material demand of a deep learning model based on swarm intelligence, and aims to solve the problems that a conventional hyper-parameter optimization method is low in efficiency and is liable to fall into local optimum. The method comprises the following steps: firstly, constructing a deep learning model containing the number of layers, the number of hidden nodes, a learning rate and regularization parameters, and defining an optimization interval of each hyper-parameter; secondly, designing an improved sparrow search algorithm (ISSA), enhancing global search capability through Cauchy variation disturbance, balancing exploration and development of a refraction reverse learning strategy, and adjusting a self-adaptive early warning value to improve convergence stability so as to optimize a hyper-parameter combination; further taking a verification set RMSE as a fitness target, combining with parallelization calculation to accelerate model performance evaluation, and screening optimal hyper-parameter configuration; and finally, predicting material demand data by using the optimized model, performing transverse comparison with LSTM, ARIMA, GRU, GCN and other models, and verifying the effectiveness of the model through multiple indexes such as RMSE, MAPE and R. Experiments show that the ISSA optimization method provided by the invention significantly reduces prediction errors, the optimized model has higher accuracy and robustness in a complex material demand scene, and meanwhile, the reliability of performance improvement is confirmed through statistical significance test.

Description

Technical Field

[0001] The present invention belongs to the field of power grid logistics, and more specifically relates to a method for predicting material demand based on a deep learning model of swarm intelligence. Background Art

[0002] Grid material demand forecasting is a crucial component of power system operation and management. Its accuracy directly impacts the safe, stable operation and economic benefits of the power system. The primary task of this forecast is to predict the quantity of materials required by the grid over a specific period of time, enabling timely procurement and dispatch. Furthermore, accurate forecasting can prevent or reduce material waste and improve resource efficiency.

[0003] Currently, power grid material demand forecasting primarily relies on statistical methods (such as time series analysis and ARIMA models) and machine learning methods (such as neural networks and support vector machines). However, these methods have several shortcomings. First, traditional statistical methods, based on historical data, may not accurately capture the real-time dynamic characteristics of the forecast. Second, while machine learning methods offer strong predictive performance, selecting appropriate model parameters (or hyperparameters) is a complex process that requires significant computational resources and time. Furthermore, existing methods are generally ineffective for forecasting power grid material demand in complex environments.

[0004] Given this, existing methods for forecasting power grid material demand often suffer from low prediction accuracy, weak model generalization, and difficulty in parameter selection. There is an urgent need to research new forecasting methods to improve the accuracy and feasibility of power grid material demand forecasting. In this paper, we propose a method for forecasting power grid material demand using a deep learning model based on swarm intelligence. This method uses an optimization algorithm to select appropriate hyperparameters, improving the model's prediction accuracy and generalization. Summary of the Invention

[0005] This paper aims to address the current technical challenges of low prediction accuracy, weak model generalization, and difficulty selecting appropriate model hyperparameters in power grid material demand forecasting. To this end, we propose a method for forecasting power grid material demand using a deep learning model based on swarm intelligence. This method combines the powerful predictive capabilities of deep learning with the global search advantages of swarm intelligence optimization algorithms to improve prediction accuracy, enhance model generalization, and effectively address the issue of model hyperparameter selection.

[0006] In order to achieve the above object, the present invention is implemented by adopting the following technical solution: comprising the following steps:

[0007] (a) Construct a deep learning model architecture including the number of layers, number of hidden nodes, learning rate, and regularization parameters, and define the optimization range of each hyperparameter;

[0008] (b) An improved sparrow search algorithm (ISSA) is used to optimize the hyperparameter combination, where the improved strategies include Cauchy mutation perturbation, position update based on reverse learning, and adaptive warning value adjustment;

[0009] (c) Using the validation set RMSE as the fitness target, the model performance of different hyperparameter combinations is evaluated in parallel;

[0010] (d) The model is trained by combining optimized hyperparameters, and predictions are made using real material demand data. The model is then compared and verified with LSTM, ARIMA, GRU, and GCN models.

[0011] In one solution, the Cauchy variation perturbation in step (b) is specifically as follows: in the process of generating the global optimal solution, the Cauchy distribution function is used to perturb the individual positions of the sparrows to escape from the local optimum.

[0012] In one embodiment, the position update of the reverse learning adopts a refractive reverse learning strategy to dynamically adjust the dimensional boundaries of the search space to balance the exploration and development capabilities.

[0013] In one solution, the adaptive warning value decreases linearly according to the number of iterations, and the proportion of alerters is controlled to improve the convergence stability of the algorithm.

[0014] In one embodiment, the parallel evaluation in step (c) synchronously trains models corresponding to multiple hyperparameter combinations through a distributed computing framework.

[0015] In one embodiment, the hyperparameter optimization range includes: the number of layers is an integer of 1-5 layers, the number of hidden nodes in a single layer is 50-500, and the learning rate and regularization parameter are logarithmically uniformly distributed from 1e-4 to 0.1.

[0016] In one solution, the horizontal indicator comparison verification includes RMSE, MAPE and R 2 indicators and confirm the performance differences through statistical significance tests.

[0017] In one embodiment, the model training adopts an early stopping strategy, and the training is terminated when the validation loss does not decrease for five consecutive rounds to prevent overfitting.

[0018] Beneficial effects of the present invention:

[0019] This paper proposes a method for predicting power grid material demand based on a deep learning model of swarm intelligence. It combines the powerful prediction capabilities of deep learning with the global search advantages of swarm intelligence optimization algorithms, and has the following beneficial effects:

[0020] High-precision forecasting: Using deep learning models and swarm intelligence optimization algorithms, compared with traditional forecasting methods, it can more accurately capture and reflect the dynamic changes and nonlinear characteristics of power grid material demand, thereby significantly improving forecast accuracy.

[0021] Strong generalization capability: Deep learning models have strong feature learning and generalization capabilities, and can effectively process complex, high-dimensional, and nonlinear data in power grid material demand forecasting, thereby improving the generalization capability of the forecasting model.

[0022] Hyperparameter optimization: Through the swarm intelligence optimization algorithm, the present invention can automatically find the optimal model hyperparameter combination without the need to manually set the hyperparameters. This not only simplifies the model training process, but also avoids the degradation of model performance due to inappropriate hyperparameter settings.

[0023] Strong practicality: The method of the present invention has high practicality and can be widely used in the actual operation of power grid enterprises, helping enterprises to make more accurate and efficient material demand forecasts. At the same time, the training and forecasting processes of the model can be automated, greatly improving the efficiency of material demand forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The flowchart of the improved ISSA algorithm of the present invention;

[0025] Figure 2 This is a horizontal comparison chart of the superiority of the improved sparrow search algorithm. DETAILED DESCRIPTION

[0026] To facilitate understanding of the present invention, the present invention will be described more fully below with reference to the accompanying drawings. The drawings illustrate exemplary embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as those understood by those skilled in the art to which the present invention pertains. The terms used in the present specification are for the purpose of describing specific embodiments only and are not intended to limit the present invention. To facilitate understanding of the present invention, a more comprehensive description of the present invention will be provided below with reference to the accompanying drawings. Typical embodiments of the present invention are shown in the drawings. However, the present invention may be embodied in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0028] Deep learning models require the selection of appropriate hyperparameters, such as the number of layers, the number of hidden nodes per layer, the learning rate, and the regularization parameter. Choosing appropriate hyperparameters can improve model performance, but manually setting hyperparameters or traditional methods like gradient descent are ineffective when there are too many hyperparameter combinations. Therefore, it is necessary to design reasonable and effective hyperparameter optimization methods. To improve the implementation of the deep learning models involved in the aforementioned research, we first determine the optimal range of hyperparameter values ​​that need to be optimized based on the structure of the designed deep learning model. Based on this, we design an optimized population set generation strategy for the model hyperparameters. Using RMSE as the optimization target, we design an improved intelligent optimization algorithm to optimize the required hyperparameter combinations for the model. Finally, we conduct model experiments using real data and conduct a horizontal comparison and evaluation with other common deep learning prediction models (such as LSTM, ARIMA, GRU, GCN, etc.). We also comprehensively analyze evaluation metrics (such as RMSE, MAPE, and R2) to compare the performance of different models to verify the effectiveness of the designed deep learning model.

[0029] The basis for constructing a deep learning model of swarm intelligence in the present invention is: the Sparrow Search Algorithm (SSA) is inspired by the foraging behavior and anti-predation behavior of sparrows. SSA benefits from its rich and characteristic search mechanism and has strong search capabilities, but when solving complex problems, it is still prone to premature maturity and falling into local optimality. By incorporating the Cauchy operator, the effect of the variation at both ends of the Cauchy distribution function is fully utilized to optimize the global optimal individual, and on the basis of reverse learning, the law of refraction of light is combined to find the optimal solution. In addition, the higher the warning value, the easier it is for sparrows to perceive the presence of natural enemies, and thus move to other areas to improve the global search capability of the algorithm. In order to improve its convergence speed, an adaptive strategy is introduced to update the warning value. The Sparrow Search Algorithm (SSA) was proposed in 2020 as a relatively novel bionic optimization algorithm.

[0030] 1) Foraging behavior

[0031] The finder in the sparrow population is responsible for searching for areas with high food density within the region, thus providing foraging areas and directions. Finders with better fitness values ​​will have limited access to food. After the followers reach the finder's foraging search range, the finder can obtain a larger search space. The finder's position is updated as follows:

[0032]

[0033] Among them, t represents the current number of iterations, iter max Indicates the maximum number of iterations. i，jrepresents the position of the i-th sparrow in the j-th dimension. α∈(0,1] represents a random number. R2∈[0,1] represents the warning value, ST∈[0.5,1] represents the safety value, Q represents a random number that follows a normal distribution, and L represents a dimension-based all-one matrix.

[0034] After a follower joins the finder's foraging area, some followers will monitor the finder. If the finder finds a better foraging area, the follower will immediately leave the current location to grab the finder's new foraging area. If the follower succeeds, it will immediately get food. If the grab fails, the follower will leave the current area due to its low fitness value and go to another finder's area. The follower's position is updated as follows:

[0035]

[0036] A + =A T (AA T ) -1 (158)

[0037] Among them, X P represents the optimal position occupied by the discoverer, X worest Represents the global worst position. A represents a matrix whose elements are assigned 1 or -1 based on the dimension.

[0038] 2) Anti-predation process

[0039] When sparrow populations are foraging, they will also be preyed upon by corresponding natural enemies. Therefore, when predators appear, sparrow populations will develop a sense of danger, and sparrow populations on the edge will move to low-risk locations, while sparrow populations in the middle will move to random locations.

[0040] The sparrow population position update during the anti-predation process is as follows:

[0041]

[0042] in, is the current global optimal value; β is the step size control parameter; K∈[-1,1]; f i is the fitness value of the current sparrow individual; f g and f w are the current global optimal and worst fitness values ​​respectively; ε is a constant.

[0043] A method for predicting material demand based on a deep learning model based on swarm intelligence Step 1. Establish a deep learning model based on swarm intelligence

[0044] The Sparrow Search Algorithm (SSA) benefits from its rich and unique search mechanism, which gives it strong global search capabilities and allows it to achieve better optimization accuracy than general swarm intelligence algorithms. However, when solving large-scale optimization problems, like most swarm intelligence algorithms, it suffers from premature algorithm maturation and a tendency to fall into local optimality. Therefore, it cannot be used directly to solve models. Therefore, based on the SSA search mechanism and targeting the established model, the following improvement strategies are introduced to enhance the algorithm's local and global search capabilities.

[0045] A. Cauchy mutation

[0046] The Cauchy distribution function has a small peak at the origin but a long distribution at both ends. Using Cauchy mutation can generate larger disturbances near the currently mutated sparrow individual, thus widening the range of the Cauchy distribution function. Cauchy mutation at both ends of the distribution makes it easier to escape the local optimum. Incorporating the Cauchy operator and fully utilizing the effects of mutation at both ends of the Cauchy distribution function to optimize the global optimal individual allows the algorithm to better reach the global optimum.

[0047] In finding the optimal solution Then, the Cauchy mutation formula is as follows:

[0048]

[0049] Among them, Cauchy(0,1) is the standard Cauchy sequence.

[0050] B. Reverse Learning

[0051] Based on reverse learning, the optimal solution is found by combining the law of light refraction. This refraction reverse learning mechanism is a superior improvement mechanism for optimizing initial assignments and candidate solutions. When followers in a sparrow population update their positions, they can effectively optimize their location area and direction. The specific formula is as follows:

[0052]

[0053] Among them, X i,j Indicates the position of the i-th individual in the current population on the j-th dimension, The reverse solution of refraction, a j and b j are the minimum and maximum values ​​of the j-th dimension in the search space.

[0054] C. Adaptive Strategy

[0055] According to the principles of SSA, the selection of warning values ​​is closely linked to the algorithm's local optimization and global search capabilities. A higher warning value makes it easier for sparrows to detect the presence of predators and migrate to other areas of space, improving the algorithm's global search capability. Conversely, a lower warning value accelerates the algorithm's convergence. Based on the above analysis, an adaptive strategy was introduced to adjust the selection of warning values, improving convergence while maintaining global search capability. The details are as follows:

[0056]

[0057] In summary, the proposed ISSA algorithm flow is as follows Figure 1 As shown:

[0058] Step 1: Initialize the population, the number of iterations, the ratio of predators to followers, and define the decision variables;

[0059] The initialization phase requires setting the size of the sparrow population N and the maximum number of iterations T. max , the proportion of discoverers P ratio (usually 20%-30%) and the proportion of alerts S ratio (usually 5%-15%). The decision variable dimension \(D\) is determined by the material demand characteristic dimension, such as the length of the time series, the number of influencing factors, etc. The initial position X of the i-th individual in the population i =[x i1 ,x i2 ,...,x iD ]Generated by uniform distribution:

[0060] x ij =a j +(b j -a j )·U(0,1),j=1,2,...,D

[0061] Where ([a j ,b j ) is the search range boundary of the j-th dimension variable, and U(0,1) is a uniformly distributed random number. At the same time, the initial value R of the adaptive warning value R is initialized. init , Cauchy mutation probability p c and reverse learning probability p r .

[0062] Step 2: Calculate the population fitness value and sort it;

[0063] Calculate the fitness value f(X i ), weighted mean square error (WMSE) is usually used as the fitness function in material demand forecasting:

[0064]

[0065] where y t For actual needs, is the model prediction value, w t is the time decay weight w t =e -λt Sort the population according to the fitness value and select the top N·P ratio Some individuals are discoverers and the rest are followers.

[0066] Step 3: Update the discoverer's location;

[0067] The finder updates the position using the following formula:

[0068]

[0069] Where α is the attenuation coefficient (usually 0.8-0.95), Q~N(0,1) is Gaussian noise, Represents the average distance from the optimal individual. The adaptive warning value R is dynamically adjusted according to the following formula:

[0070]

[0071] Where γ controls the decay rate (usually 1.5-2.5).

[0072] Step 4: The follower monitors the discoverer all the time. If it notices that the discoverer has found better food, it updates the follower's position.

[0073] Followers update rules based on the discoverer's location:

[0074]

[0075] where X p is a randomly selected discoverer individual, β~U(0.5,1.5) is the follow-up coefficient, It is a dynamic learning rate to ensure that the later stages of iteration tend to be refined searches.

[0076] Step 5: Randomly select a certain proportion of alerts and update their positions according to the public information;

[0077] Randomly select N·S ratio Individuals perform vigilance behaviors:

[0078]

[0079] in Control the disturbance intensity, use the tangent function to generate long-tail disturbances, and enhance the ability to escape from local optimality.

[0080] Step 6: Select the Cauchy mutation stirring strategy and reverse learning strategy according to the probability to stir the current optimal solution and generate a new solution.

[0081] For the current optimal solution X best′ Apply a mixed perturbation strategy:

[0082] 1. Cauchy mutation (probability p c ):

[0083]

[0084] 2. Refraction reverse learning (probability p r ):

[0085]

[0086] in Indicates element-wise multiplication, κ is the refractive index (usually 0.5-1.2). By greedy selection, retain argmin{f(X best ),f(X bes t′),f(X ROBL )} as the new optimal solution.

[0087] Step 7: Calculate fitness and update sparrow position;

[0088] Recalculate the fitness values ​​of all individuals and update the population according to the elite retention strategy: retain the top 10% of the best individuals and directly enter the next generation, and the remaining 90% will be selected through tournament selection. For the material demand forecasting task, the deep learning model parameters θ need to be updated simultaneously:

[0089]

[0090] Implement hybrid training of swarm intelligence optimization and gradient descent.

[0091] Step 8: If the maximum number of iterations is reached and the output conditions are met, the optimal solution and fitness value are output; otherwise, return to Step 2.

[0092] When t≥T max Or the change in optimal fitness for K consecutive generations (Threshold ò=10 -6 ) when the algorithm is terminated and the optimal solution X is output best and its corresponding prediction model parameters. Otherwise, return to Step 2 to continue iterating and update the adaptive parameters at the end of each generation:

[0093]

[0094] First, the population must be initialized to determine the size of the sparrow population, the maximum iteration threshold, the initial ratio of discoverers to followers, and the dimensionality of the decision variables and the scope of the search space. The position vector of each individual sparrow in the population is generated within the pre-set variable space using uniform random sampling. The fitness function is then defined based on the characteristics of the problem. For example, in a material demand forecasting scenario, mean squared error can be used as an evaluation metric. During the initialization phase, improvement strategy parameters such as the Cauchy mutation probability coefficient and the reverse learning trigger threshold must be configured, and initial weight parameters must be assigned to the adaptive warning values.

[0095] After initialization, the iterative optimization phase begins. The fitness values ​​of all individuals in the current population are calculated and ranked according to fitness to determine the current optimal individual. The discoverer's position is updated according to the position movement formula of the standard sparrow search algorithm. Its exploration step size is dynamically adjusted by an adaptive warning value, which decays with the number of iterations through a nonlinear function. This maintains a large step size in the early stages of the iteration to enhance global search capabilities, while reducing the step size in the later stages to improve convergence accuracy. Followers continuously monitor the discoverer's position changes. When they detect that the discoverer's position has superior fitness, they adjust their own position using the Lévy flight strategy. This weighted offset is performed by combining the current optimal individual position with the population average position. A Gaussian perturbation term is introduced during the displacement process to prevent path rigidity.

[0096] The update mechanism for the vigilant randomly selects a specific proportion of individuals in the population to perform vigilance behavior, dynamically adjusting the vigilance range based on the current number of iterations and spatial dimensions. While performing a localized, refined search near the optimal individual, it also jumps to the boundary of the search space with a certain probability to detect potential optimal solutions. The collaborative application of Cauchy mutation and reverse learning strategies occurs during the elite solution optimization phase of each iteration. The perturbation strategy is selected based on a preset probability distribution: when performing Cauchy mutation, a random perturbation based on the standard Cauchy distribution is applied to the current optimal solution, enhancing the ability to escape the local optimum through the long-tail nature of the probability density function; when using reverse learning, a refracted inverse solution of the current solution is constructed according to the refraction law, leveraging spatial symmetry to expand the search area. The refraction angle is determined by the quality of the solution and the boundary of the search space, and the reverse mapping value of each variable is independently calculated in the dimension.

[0097] After each policy perturbation, the fitness of the new solution is recalculated, and a greedy selection mechanism is used to determine whether the new solution is accepted as the current optimal solution. After the population position is updated, a check is performed to determine whether the preset maximum number of iterations has been reached or whether the fitness convergence threshold (e.g., the change in the optimal solution over five consecutive generations is less than 1e-6) has been met. If the termination criteria are met, the position vector of the individual with the optimal fitness is output as the final solution; otherwise, the next iteration cycle is entered. Throughout this process, a dynamic adjustment mechanism for the adaptive early warning value is continuously in effect, with its attenuation coefficient feedback-adjusted based on the population diversity indicator. When the population standard deviation is detected to be below a critical value, the mutation probability is automatically increased, thus forming a self-regulating optimization closed loop. The entire implementation process requires particular attention to computational efficiency in high-dimensional spaces. This can be accelerated through techniques such as dimensional grouping and parallel computing, and by prioritizing important variables.

[0098] Step 2. Solve the established deep learning model of swarm intelligence.

[0099] Benchmark function test related information

[0100]

[0101]

[0102] The global optimization ability and effectiveness of the algorithm are crucial to solving the established model. Therefore, the steps for verifying the basic test functions of the improved algorithm are further expanded as follows: First, to verify the solving performance of the ISSA algorithm, four groups of typical CEC basic test functions are selected to evaluate the algorithm's exploration ability, stability and global search ability. In the comparative experiment, the particle swarm algorithm (PSO), genetic algorithm (GA) and sparrow search algorithm (SSA) are selected for comparison.

[0103] When solving the established swarm intelligence deep learning model, the first step is to ensure that the improved sparrow search algorithm (ISSA) possesses sufficient global optimization capabilities and stability. To this end, validating the algorithm's effectiveness through benchmarking functions is a key step. The specific implementation process is as follows: Four sets of classic CEC benchmark functions are selected, covering different types, including unimodal, multimodal, high-dimensional, and fixed-dimensional. For example, the Sphere function verifies the algorithm's convergence accuracy and speed, the GeneralizedPenalized function tests global search capabilities, the Quartic Function assesses robustness against noise interference, and the Goldstein-Price function verifies its ability to search complex terrain in low-dimensional space. These functions, through their different characteristics, simulate the difficulties of real-world optimization problems, such as local extrema traps, the curse of dimensionality, and nonlinear noise. In comparative experiments, the ISSA is compared with the particle swarm optimization algorithm (PSO), the genetic algorithm (GA), and the original sparrow search algorithm (SSA) to ensure comprehensiveness. All algorithms are set to the same parameters: population size, maximum number of iterations, and search space range to reduce external interference. The experiment was run independently 30 times to eliminate the influence of randomness. The optimal solution, average solution, and standard deviation that the algorithm converged to were recorded in each run to quantify the stability of the algorithm. For example, when solving the Sphere function, ISSA can quickly jump out of the local optimum through the synergistic effect of Cauchy mutation and refractive reverse learning, and always approaches the theoretical optimal value of 0 in multiple experiments, while PSO and SSA have large deviations due to premature convergence; in the Goldstein-Price function, ISSA accurately locates the global optimum with its adaptive early warning strategy, while other algorithms fall into suboptimal solutions due to their inability to balance exploration and development. Finally, by statistically analyzing the average error and standard deviation of each algorithm on different test functions, the advantages of ISSA in convergence speed, global search robustness, and anti-interference ability are verified, thereby proving its suitability for parameter optimization of deep learning models. This process not only provides algorithm performance guarantees for model solving, but also lays a theoretical foundation for the global optimization of model parameters in subsequent material demand forecasting tasks. The results are as follows Figure 2 shown.

[0104] Example 1:

[0105] The specific implementation process of a provincial power grid company using this method to forecast power grid material demand is as follows:

[0106] To address the impact of complex weather and renewable energy integration on grid material dispatch, the province constructed a forecasting model using historical data from 2018 to 2022. This data covers the quarterly consumption of 28 core materials, including transformers, high-voltage cables, insulators, and smart meters, across 16 prefecture-level cities across the province. The model also incorporates meteorological data such as temperature, rainfall, and wind speed, as well as 12 economic indicators, including regional GDP, industrial electricity consumption, and installed renewable energy capacity. Furthermore, the model includes records of emergency material allocations during 32 natural disasters, including typhoons and wildfires.

[0107] During the model construction phase, historical demand data and external features were first aligned in time and space to generate an input matrix containing 486 feature dimensions. The improved sparrow search algorithm (ISSA) was used to optimize the bidirectional LSTM network structure, with an initial population size of 120 sparrow individuals. Each individual encoded the number of hidden layer neurons in the LSTM (ranging from 80 to 200), the time window length (4 to 12 quarters), and the drop rate (0.1 to 0.3). Using the Cauchy mutation mechanism, the optimal individual was perturbed at the 50th iteration of the algorithm, successfully breaking through the local optimal trap caused by the extreme cold wave data in the winter of 2020 and increasing the prediction model's sensitivity to abnormal weather by 19%.

[0108] During implementation, when the algorithm detected sustained high temperatures exceeding 40°C in southern Jiangsu during the third quarter of 2023, the adaptive early warning module dynamically adjusted the warning threshold from an initial value of 0.8 to 0.65, triggering a reverse learning mechanism to generate a reverse solution for material demand. Combined with real-time monitoring data showing distribution transformer load rates exceeding 85% in Changzhou and Wuxi, the model predicted that demand for 630kVA oil-immersed transformers in the region would surge from 120 units in the regular quarter to 178 units, with a discrepancy of only 3.5% with the actual procurement of 172 units. For emergency supplies forecasts before and after Typhoon Meihua, the model used reverse learning to deduce the sudden inflection point in insulator demand, generating a 72-hour advance warning of the need for 230,000 additional disc suspension insulators in Qingdao and Yantai, improving forecast timeliness by 41% compared to traditional ARIMA models.

[0109] After 12 months of operational validation, this method achieved a mean absolute percentage error (MAPE) of 4.2% in conventional material forecasting, a 7.8 percentage point reduction compared to the unmodified SSA-LSTM model. In response to a sudden wildfire in northern Henan Province in the summer of 2023, it accurately predicted the specifications and quantities of materials such as steel-core aluminum stranded wire and tension clamps required for a 12-kilometer 35kV overhead line, shortening the time it takes to deploy emergency supplies from the typical 48 hours to 29 hours. Calculations show that this new method has increased the provincial power grid's material inventory turnover rate from an average of 3.2 times per year to 4.5 times, reduced the proportion of obsolete inventory from 12.7% to 6.9%, and saved 23 million yuan in annual material management costs.

[0110] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0111] It should be understood that the detailed description of the technical solutions of the present invention using the preferred embodiments above is illustrative and not restrictive. A person skilled in the art, after reading the present specification, may modify the technical solutions described in the embodiments or replace some of the technical features therein with equivalents; such modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting material demand based on a deep learning model of swarm intelligence, characterized in that: The following steps are involved: (a) Construct a deep learning model architecture including the number of layers, number of hidden nodes, learning rate, and regularization parameters, and define the optimization range of each hyperparameter; (b) An improved sparrow search algorithm (ISSA) is used to optimize the hyperparameter combination. The improved strategies include Cauchy mutation perturbation, position update based on reverse learning, and adaptive warning value adjustment. (c) Using the validation set RMSE as the fitness target, the model performance of different hyperparameter combinations is evaluated in parallel; (d) The model is trained by combining optimized hyperparameters, and predictions are made using real material demand data. The model is then compared and verified with LSTM, ARIMA, GRU, and GCN models.

2. The method according to claim 1, wherein The Cauchy variation perturbation in step (b) is specifically as follows: in the process of generating the global optimal solution, the Cauchy distribution function is used to perturb the individual positions of the sparrows to escape from the local optimum.

3. The method according to claim 1, wherein The position update of the reverse learning adopts a refractive reverse learning strategy, which dynamically adjusts the dimensional boundary of the search space to balance the exploration and development capabilities.

4. The method according to claim 1, wherein The adaptive warning value decreases linearly according to the number of iterations, and the proportion of alerters is controlled to improve the convergence stability of the algorithm.

5. The method according to claim 1, wherein The parallel evaluation in step (c) synchronously trains models corresponding to multiple hyperparameter combinations through a distributed computing framework.

6. The method according to claim 1, wherein The hyperparameter optimization range includes: the number of layers is an integer of 1-5 layers, the number of hidden nodes in a single layer is 50-500, and the learning rate and regularization parameter are logarithmically uniformly distributed from 1e-4 to 0.

1.

7. The method according to claim 1, wherein The horizontal indicator comparison verification includes RMSE, MAPE and R² indicators, and the performance differences are confirmed through statistical significance tests.

8. The method according to claim 1, wherein The model training adopts an early stopping strategy, and the training is terminated when the validation loss does not decrease for five consecutive rounds to prevent overfitting.

Images