Internet of Things time series data prediction method and system based on intelligent group optimization deep learning

Abstract

The invention relates to an Internet of Things time series data prediction method and system based on intelligent group optimization deep learning, and belongs to the technical field of Internet of Things. The method comprises the following steps: standardizing time series data of the Internet of Things; decomposing the standardized data into high-frequency, intermediate-frequency and low-frequency component groups by adopting an empirical mode decomposition method; an improved bat algorithm is constructed, wherein the improvement comprises the steps of introducing a dynamically changing inertia weight, adopting a Levy flight strategy, introducing Cauchy random disturbance and adopting linear progressive decrease adjustment on loudness and pulse emissivity; performing adaptive optimization on hyper-parameters of a deep learning prediction model by using the improved bat algorithm; and finally, training independent prediction models for the high-frequency component group, the medium-frequency component group and the low-frequency component group by using the optimal hyper-parameter combination, performing linear superposition reconstruction on the prediction results of the components, and executing a destandardization operation to obtain a final prediction value. According to the method, the accuracy of time series data prediction of the Internet of Things and the optimization efficiency of the model are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) technology and relates to a method and system for predicting IoT time-series data based on intelligent swarm optimization deep learning. Background Technology

[0002] With the rapid development of IoT technology, massive numbers of sensors and smart terminal devices are widely deployed in various monitoring scenarios, generating large-scale time-series data in real time. Efficient analysis of this complex time-series data, especially anomaly detection, is crucial for ensuring the safe and stable operation of IoT systems and for timely detection of equipment failures or environmental anomalies.

[0003] In existing technologies, artificial intelligence methods, especially deep learning models, have become the mainstream technology for processing IoT time-series data. For example, Long Short-Term Memory networks and their variants are widely used due to their advantages in capturing time-series dependencies.

[0004] However, it is well known to those skilled in the art that the performance of deep learning models largely depends on the configuration of their hyperparameters. Taking the Long Short-Term Memory (LSTM) network model as an example, the selection of hyperparameters such as the learning rate, the number of iterations, and the number of neurons in the hidden layer has a significant impact on the prediction accuracy and convergence speed of the final model. Currently, the hyperparameter tuning process for such models often relies on tedious manual trial and error or grid search. These traditional tuning methods are not only inefficient and highly subjective, but also fail to guarantee that the searched parameter combinations are globally optimal, resulting in poor accuracy and stability of the constructed detection models, making it difficult to meet the high-precision analysis requirements of complex IoT scenarios. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method and system for predicting IoT time series data based on intelligent swarm optimization deep learning, thereby overcoming the shortcomings of existing technologies in processing IoT time series data, where the tuning of hyperparameters of deep learning models relies on human experience, has low optimization efficiency, and is difficult to guarantee optimal configuration, resulting in insufficient anomaly detection accuracy and model reliability.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting IoT time-series data based on intelligent swarm optimization deep learning, the method specifically includes the following steps: S1. Obtain the IoT time series data to be analyzed, and use the minimum-maximum standardization method to constrain the data to the [0,1] interval; use empirical mode decomposition to decompose the data into several intrinsic mode functions, and divide them into high-frequency, mid-frequency and low-frequency component groups according to frequency characteristics; S2. To address the shortcomings of the standard bat algorithm, a multi-strategy improvement is implemented: a dynamic inertial weight that is adaptively adjusted with the number of iterations is introduced; a Levy flight strategy is adopted to enhance the global search; a Cauchy random perturbation is introduced to enhance the local search; and loudness and pulse emission rate are adjusted linearly and gradually. S3. Adaptively optimize the key hyperparameters of the deep learning prediction model using the improved bat algorithm; use the model's performance index on the validation set as the fitness function, and iteratively search to output the globally optimal hyperparameter combination. S4. Using the optimal hyperparameter combination obtained through optimization, construct and train independent deep learning prediction models for the high-frequency component group, mid-frequency component group and low-frequency component group obtained in S1 respectively, to capture data features at different frequency scales. S5. The prediction results of each independent model are linearly superimposed and reconstructed. The merged data is then de-standardized to obtain the final prediction value with actual physical meaning. The effectiveness of the prediction model is quantitatively evaluated using indicators such as root mean square error.

[0007] Furthermore, in step S1, the specific method for dividing the decomposed components into high-frequency component groups, mid-frequency component groups, and low-frequency component groups is as follows: High-frequency component group: It consists of several initial intrinsic mode function components, used to characterize random noise or transient change information in the data; Intermediate frequency component group: Composed of several intermediate intrinsic mode function components, used to reflect the main periodic fluctuations or key event information in the data; Low-frequency component group: Composed of the remaining intrinsic mode function components and residual terms, used to represent the overall trend or long-term variation law of the original time series data.

[0008] Furthermore, in step S2, the improvement to the standard bat algorithm speed update formula refers to the introduction of a factor that varies with the number of iterations. Adaptive Adjustment of Nonlinear Dynamic Inertia Weights The calculation formula is as follows: (1) in, This is the preset maximum value for the inertia weight. This is the preset minimum value for the inertia weight. This represents the current iteration number. The maximum number of iterations is denoted by . This dynamic inertia weight gives the algorithm a large value in the early stages of iteration, thus giving the individual bats a greater speed to enhance their global search ability and quickly explore the entire solution space. In the later stages of iteration, the value of the inertia weight gradually decreases, thereby reducing the speed of the individual bats and enabling them to perform fine-grained searches within the neighborhood of the discovered optimal solution. This effectively balances the algorithm's global exploration and local development capabilities, significantly improving the algorithm's convergence speed and optimization accuracy.

[0009] Furthermore, in step S2, to enhance the algorithm's global search capability, the Levy flight strategy is adopted to replace the inherent random flight in the standard bat algorithm. This addresses the problems of decreased population diversity and susceptibility to local optima caused by the endogeneity of the search method in the standard algorithm. The improved bat individual position update formula is as follows: (2) in, For individuals In the Position at the next iteration This is the step size control coefficient. For dot product operation, and The random step size is generated based on the Lévy distribution. This strategy combines the characteristics of small-scale fine-grained search and large-scale random jumps when organisms are foraging, allowing individual bats to explore locally around the current solution for most of the time, while also making long-distance jumps with a certain probability. This effectively expands the search range of the population, increases population diversity, and enables the algorithm to jump out of the attraction domain of the current local optimum with a high probability when the optimization stalls, significantly enhancing the algorithm's global exploration capability.

[0010] Furthermore, in step S2, to enhance the algorithm's local search capability, a Cauchy random perturbation is introduced instead of the uniform random perturbation in the standard algorithm when the individual bat performs a local search. This overcomes the problem that the standard algorithm is inefficient in searching near the optimal solution and is prone to getting stuck in local optima. The improved local search position update formula is as follows: (3) in, This represents the optimal position of an individual in the current population. For disturbance control factors, Representing the standard Cauchy distribution, this strategy utilizes the long tails at both ends of the Cauchy distribution, allowing newly generated positions to move away from the current optimal position with a certain probability. Compared to uniform perturbation with a fixed range, Cauchy perturbation can generate a larger range of random step sizes, thereby giving individuals stronger perturbation capabilities. When trapped in a local optimum, it can jump out of the region with a greater probability, significantly improving the algorithm's local exploration vitality and final optimization accuracy.

[0011] Furthermore, the improvement made in step S2 to accelerate the later convergence of the algorithm lies in addressing the loudness issue in the standard bat algorithm. and pulse emission rate To address the issue of decreased optimization capability in later stages of the algorithm when using a fixed step size update, a dynamic adjustment method with linear decreasing steps is adopted. The update formula is as follows: (4) (5) In the formula, and These are the preset minimum and maximum values ​​for loudness, respectively; and These are the preset minimum and maximum values ​​for the pulse emission rate, respectively; This represents the current iteration number; This is the maximum number of iterations. This adjustment strategy makes the loudness... The pulse emission rate gradually decreases as iterations proceed, while the pulse emission rate... The loudness gradually increases, thus achieving a smooth transition of the algorithm's behavior from global exploration to local development. In the early stages of the algorithm, a larger loudness and a smaller pulse emission rate help to accept new solutions and encourage global search. In the later stages of the algorithm, a smaller loudness and a larger pulse emission rate make the algorithm more inclined to conduct fine development near the current optimal solution. This effectively accelerates the convergence speed of the algorithm in the later stages.

[0012] Furthermore, in step S3, the process of adaptively optimizing the key hyperparameters of the deep learning prediction model using the improved bat algorithm specifically involves: using the performance evaluation index of the deep learning model on a preset validation set as the fitness function of the improved bat algorithm, and encoding the combination of hyperparameters to be optimized as the position of the individual bat; through iterative optimization, finding the position of the individual bat that optimizes the fitness function, thereby determining the optimal combination of hyperparameters for the deep learning model.

[0013] Furthermore, in step S4, the deep learning prediction model is specifically a long short-term memory network model or a variant thereof, including but not limited to gated recurrent unit networks or bidirectional long short-term memory networks; such models are suitable for processing prediction tasks of IoT time series data because they have the ability to capture long-term dependencies in time series data.

[0014] The present invention also provides an IoT time series data prediction system based on intelligent swarm optimization deep learning.

[0015] The beneficial effects of this invention are as follows: Compared with existing technologies, this invention has the following significant advantages through multifaceted improvements to the standard bat algorithm: 1. By introducing dynamic inertia weights This allows the algorithm to have a high speed in the early stages of iteration to support global search, while reducing the speed in the later stages of iteration to ensure fine search in the neighborhood of the optimal solution, which significantly improves the optimization efficiency and accuracy in the later stages of iteration.

[0016] 2. The Levy flight strategy is used to replace the endogeneity-based global search formula, which effectively expands the search range of the population, increases population diversity, and avoids the algorithm from getting trapped in local optima too early.

[0017] 3. The introduced Cauchy random perturbation is more flexible than the uniform random perturbation, enhancing the perturbation ability of individuals in their local neighborhoods, making it difficult for the algorithm to converge to a local optimum and easier to break through local extrema.

[0018] 4. Adjust loudness using a linear decrease method. and pulse emission rate This accelerates the convergence of the algorithm in the later stages and ensures that the algorithm can smoothly transition from global search to local exploration in the middle of the iteration.

[0019] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 A method flowchart is provided for this invention; Figure 2 This is a flowchart of the improved bat algorithm of the present invention. Detailed Implementation

[0021] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0022] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0023] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0024] Figure 1 A method flowchart is provided for this invention, such as Figure 1 As shown, the IoT time-series data prediction method based on intelligent swarm optimization deep learning of the present invention includes the following steps: Step S1: Data Acquisition and Preprocessing. Acquire the IoT time-series data to be analyzed and perform preprocessing operations such as data cleaning and normalization.

[0025] In IoT applications, the raw sensor data collected, such as electricity load data, is typically of a large magnitude. Furthermore, data from different sensors or monitoring indicators may belong to different dimensions. Directly feeding such raw data with different ranges and dimensions into the prediction model for training will reduce the model's convergence speed and prediction accuracy.

[0026] Therefore, the embodiments of the present invention need to standardize the collected raw data to prevent data of different ranges and dimensions from adversely affecting the performance of the model's learning features.

[0027] In this embodiment, a minimum-maximum standardization method is used. This method can constrain the range of data to... Within the range.

[0028] Specifically, the standardized calculation is shown in the following formula:

[0029] In the formula: This is a sample of raw, unprocessed data. and These are the minimum and maximum values ​​in the data sample set, respectively. This is data obtained after standardization.

[0030] By performing this step, the range of data values ​​is effectively shortened, which not only helps to shorten the training time of subsequent models, but also further improves the accuracy of model predictions.

[0031] Accordingly, the subsequent model outputs the prediction results. Afterwards, the obtained prediction results must be destandardized in order to restore the data to the normal numerical range.

[0032] The calculation of the inverse standardization is shown in the formula:

[0033] In the formula: This is the final predicted value with practical physical meaning obtained after destandardization; The standardized predicted values ​​output by the model.

[0034] Step S2: After standardizing the time series data, in order to further analyze the compositional characteristics of different frequency scales within the data, this embodiment of the invention employs the empirical mode decomposition method to process the standardized data obtained in step S1. Decompose it.

[0035] Empirical Mode Decomposition (EMD) is an adaptive signal processing technique that decomposes complex non-stationary time-series data into several intrinsic mode function (EMF) components and a residual term. Each EMF component represents an oscillation mode with different frequencies and amplitudes in the original data.

[0036] In this embodiment, after processing by the empirical mode decomposition method, the original time series data is decomposed into N intrinsic mode function components and M residual term components.

[0037] To facilitate subsequent modeling and analysis of different frequency characteristics, this invention divides the total N+M components obtained from the decomposition into three groups: high, medium, and low frequencies, based on their frequency characteristics. High-frequency components: These components have high frequencies and fluctuate dramatically in amplitude, and usually represent random noise or transient changes in the data.

[0038] Mid-frequency components: These components have moderate frequencies and typically reflect the main periodic fluctuations or key event information in the data.

[0039] Low-frequency components and residuals: These components have low frequencies and gradual changes, and usually represent the overall trend or long-term variation pattern of the original time series data.

[0040] Through the aforementioned empirical mode decomposition steps, the complex original time-series data is transformed into multiple subsequences with different scales and characteristics. Subsequent prediction models can be trained and predicted separately for these three sets of components with different characteristics. Finally, the prediction results of each set are superimposed to more precisely capture the inherent patterns of the data and improve the overall prediction accuracy.

[0041] Step S3: To address the common problems in the basic bat algorithm, improvements and adjustments are made to the basic bat algorithm in several aspects, including encoding methods and search strategies. Figure 2 This is a flowchart of the improved bat algorithm of the present invention.

[0042] Step S3.1: For the standard bat algorithm, the speed coefficient is kept constant. This leads to a decline in optimization capability in later stages. The embodiments of this invention introduce dynamically changing inertia weights. To optimize the speed formula.

[0043] This dynamically adjusted inertial weight This allows the algorithm to continuously adjust its search speed as it iterates. In the initial global search phase, the algorithm needs a high speed to quickly traverse the entire space; while in the later, more detailed search phase, a lower speed is required to ensure a more detailed search within the neighborhood where the optimal solution has been largely identified.

[0044] The improved velocity formula is shown below:

[0045] Among them, the dynamically changing inertia weight The calculation formula is as follows:

[0046] In the formula: Inertial weights; This represents the minimum value of the inertia weight. This represents the maximum value of the inertia weight. This represents the current iteration number; This represents the maximum number of iterations. This represents the frequency of the current iteration.

[0047] Step S3.2: The standard bat algorithm's global search suffers from endogeneity, leading to decreased population diversity and a tendency to get trapped in local optima. This embodiment of the invention employs the Levy flight strategy instead of random flight to enhance global search capabilities.

[0048] Lévy flight, mathematically classified as a Markov process, employs a random walk model, simulating the instinct of biological groups foraging in unknown environments. This strategy allows individuals to quickly adjust their search direction and speed when prolonged optimization efforts fail.

[0049] Therefore, this invention uses the following formula instead of the original position update formula:

[0050] Among them, Levi's flight stride The calculation is shown in the following formula:

[0051] In the formula: and The first Only one bat Second and third The position at the next iteration; This is the current optimal position; This is the step size control coefficient, which takes the value of [value] in this embodiment. ; It is a dot product; The step size is denoted as , and in this embodiment, it is taken as . ; The step size is random.

[0052] The random step size It follows the Lévy distribution and is generated using the Mantegna algorithm. Its distribution is shown in the following formula:

[0053] Step length The calculation is shown in the following formula.

[0054] In the formula, Step size, For characteristic index, and It follows a normal distribution, that is... , . and The calculation is shown in the following formula:

[0055]

[0056] In the formula, For the Gamma function, In this embodiment, the value is taken as .

[0057] Step S3.3: To address the problem that the standard BA algorithm is inefficient and prone to getting trapped in local optima when searching near the optimal solution, this embodiment of the invention introduces Cauchy perturbation instead of random uniform perturbation to enhance the individual's search ability within its local neighborhood.

[0058] Cauchy perturbation, or random perturbation according to the Cauchy distribution, generates a wider range of random numbers and is more flexible than uniform random perturbation.

[0059] The improved local search position update formula is shown below:

[0060] In the formula: The disturbance control factor is set to a value of [value] in this embodiment. ; It is a dot product; This represents the standard Cauchy distribution.

[0061] Step S3.4: Targeting loudness in the standard bat algorithm and pulse emission rate Update step size , The fixed nature of the soundness leads to poor optimization capabilities in the later stages of iteration. To accelerate convergence in the later stages, this embodiment of the invention addresses this issue by adjusting the loudness... and pulse emission rate Adjustments are made using a linear decreasing method.

[0062] This adjustment method can be applied to the number of iterations. At that time, and All are located at the midpoint of their interval range, ensuring that the algorithm transitions from global search to local exploration in the middle of the iteration.

[0063] Improved loudness The update is shown in the following formula:

[0064] Improved pulse emission rate The update is shown in the following formula:

[0065] In the formula, and These are the minimum and maximum values ​​of loudness, respectively; and These are the minimum and maximum values ​​of the pulse emission rate, respectively. This represents the current iteration number; This represents the maximum number of iterations.

[0066] Step S4: In this embodiment of the invention, the improved bat algorithm constructed in step S3 (S3.1-S3.4) is used to adaptively iteratively optimize the key hyperparameters of the subsequent long short-term memory network model.

[0067] The hyperparameters include at least the number of hidden layer neurons, the maximum number of iterations, and the learning rate. During the optimization process, the model's performance on the validation set is used as the fitness function. An improved bat algorithm is used for iterative search until the termination condition is met, outputting the optimal combination of hyperparameters.

[0068] During the model training phase, using this optimal combination of hyperparameters, multiple independent long short-term memory network prediction models are constructed and trained for the high-frequency component group, mid-frequency component group, and low-frequency component group obtained from the empirical mode decomposition in step S2, respectively.

[0069] In the model application phase, the same empirical mode decomposition is first performed on the new test data, and the resulting high, medium, and low frequency components are input into the corresponding trained LSTM-H, LSTM-M, and LSTM-L models, respectively, to obtain the predicted values ​​of each component. , , .

[0070] Subsequently, the prediction results of each component are linearly superimposed. The combined predicted value is obtained by reconstruction. .

[0071] Finally, the combined forecast values Perform the denormalization operation described in step S1, i.e. This yields the final predicted value with practical physical meaning. .

[0072] To verify the effectiveness of the method of this invention, various performance evaluation indicators can be used, such as root mean square error, mean absolute percentage error, and coefficient of determination, to evaluate the final predicted value. Compared with the true value The differences between them were quantitatively assessed.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting IoT time-series data based on intelligent swarm optimization deep learning, characterized in that, The method specifically includes the following steps: S1. Obtain the IoT time series data to be analyzed, and use the minimum-maximum standardization method to constrain the data to the [0,1] interval; use empirical mode decomposition to decompose the data into several intrinsic mode functions, and divide them into high-frequency, mid-frequency and low-frequency component groups according to frequency characteristics; S2. To address the shortcomings of the standard bat algorithm, a multi-strategy improvement is implemented: a dynamic inertial weight that is adaptively adjusted with the number of iterations is introduced; a Levy flight strategy is adopted to enhance the global search; a Cauchy random perturbation is introduced to enhance the local search; and loudness and pulse emission rate are adjusted linearly and gradually. S3. Adaptively optimize the key hyperparameters of the deep learning prediction model using the improved bat algorithm; use the model's performance index on the validation set as the fitness function, and iteratively search to output the globally optimal hyperparameter combination. S4. Using the optimal hyperparameter combination obtained through optimization, construct and train independent deep learning prediction models for the high-frequency component group, mid-frequency component group and low-frequency component group obtained in S1 respectively, to capture data features at different frequency scales. S5. The prediction results of each independent model are linearly superimposed and reconstructed. The merged data is then de-standardized to obtain the final prediction value with actual physical meaning. The effectiveness of the prediction model is quantitatively evaluated using indicators such as root mean square error.

2. The IoT time-series data prediction method based on intelligent swarm optimization deep learning according to claim 1, characterized in that, In step S1, the specific method for dividing the decomposed components into high-frequency component groups, mid-frequency component groups, and low-frequency component groups is as follows: High-frequency component group: It consists of several initial intrinsic mode function components, used to characterize random noise or transient change information in the data; Intermediate frequency component group: Composed of several intermediate intrinsic mode function components, used to reflect the main periodic fluctuations or key event information in the data; Low-frequency component group: Composed of the remaining intrinsic mode function components and residual terms, used to represent the overall trend or long-term variation law of the original time series data.

3. The IoT time-series data prediction method based on intelligent swarm optimization deep learning according to claim 2, characterized in that, In step S2, the improvement to the standard bat algorithm speed update formula refers to the introduction of a factor that varies with the number of iterations. Adaptive Adjustment of Nonlinear Dynamic Inertia Weights The calculation formula is as follows: （1） in, This is the preset maximum value for the inertia weight. This is the preset minimum value for the inertia weight. This represents the current iteration number. The maximum number of iterations is denoted by . This dynamic inertia weight gives the algorithm a large value in the early stages of iteration, thus giving the individual bats a greater speed to enhance their global search ability and quickly explore the entire solution space. In the later stages of iteration, the value of the inertia weight gradually decreases, thereby reducing the speed of the individual bats and enabling them to perform fine-grained searches within the neighborhood of the discovered optimal solution. This effectively balances the algorithm's global exploration and local development capabilities, significantly improving the algorithm's convergence speed and optimization accuracy.

4. The IoT time-series data prediction method based on intelligent swarm optimization deep learning according to claim 3, characterized in that, In step S2, to enhance the algorithm's global search capability, the Levy flight strategy is adopted to replace the random flight inherent in the standard bat algorithm. This addresses the problems of decreased population diversity and susceptibility to local optima caused by the endogeneity of the search method in the standard algorithm. The improved bat individual position update formula is as follows: （2） in, For individuals In the Position at the next iteration This is the step size control coefficient. For dot product operation, and The random step size is generated based on the Lévy distribution. This strategy combines the characteristics of small-scale fine-grained search and large-scale random jumps when organisms are foraging, allowing individual bats to explore locally around the current solution for most of the time, while also making long-distance jumps with a certain probability. This effectively expands the search range of the population, increases population diversity, and enables the algorithm to jump out of the attraction domain of the current local optimum with a high probability when the optimization stalls, significantly enhancing the algorithm's global exploration capability.

5. The IoT time-series data prediction method based on intelligent swarm optimization deep learning according to claim 4, characterized in that, The improvement in step S2 to enhance the algorithm's local search capability involves introducing a Cauchy random perturbation instead of the uniform random perturbation in the standard algorithm when an individual bat performs a local search. This overcomes the standard algorithm's low search efficiency near the optimal solution and its tendency to stagnate at local optima. The improved local search position update formula is as follows: （3） in, This represents the optimal position of an individual in the current population. For disturbance control factors, Representing the standard Cauchy distribution, this strategy utilizes the long tails at both ends of the Cauchy distribution, allowing newly generated positions to move away from the current optimal position with a certain probability. Compared to uniform perturbation with a fixed range, Cauchy perturbation can generate a larger range of random step sizes, thereby giving individuals stronger perturbation capabilities. When trapped in a local optimum, it can jump out of the region with a greater probability, significantly improving the algorithm's local exploration vitality and final optimization accuracy.

6. The IoT time-series data prediction method based on intelligent swarm optimization deep learning according to claim 5, characterized in that, The improvement made in step S2 to accelerate the later convergence of the algorithm lies in addressing the loudness issue in the standard bat algorithm. and pulse emission rate To address the issue of decreased optimization capability in later stages of the algorithm when using a fixed step size update, a dynamic adjustment method with linear decreasing steps is adopted. The update formula is as follows: （4） （5） In the formula, and These are the preset minimum and maximum values ​​for loudness, respectively; and These are the preset minimum and maximum values ​​for the pulse emission rate, respectively; This represents the current iteration number; This is the maximum number of iterations. This adjustment strategy makes the loudness... The pulse emission rate gradually decreases as iterations proceed, while the pulse emission rate... The loudness gradually increases, thus achieving a smooth transition of the algorithm's behavior from global exploration to local development. In the early stages of the algorithm, a larger loudness and a smaller pulse emission rate help to accept new solutions and encourage global search. In the later stages of the algorithm, a smaller loudness and a larger pulse emission rate make the algorithm more inclined to conduct fine development near the current optimal solution. This effectively accelerates the convergence speed of the algorithm in the later stages.

7. The IoT time-series data prediction method based on intelligent swarm optimization deep learning according to claim 6, characterized in that, In step S3, the process of adaptively optimizing the key hyperparameters of the deep learning prediction model using the improved bat algorithm is as follows: the performance evaluation index of the deep learning model on a preset validation set is used as the fitness function of the improved bat algorithm, and the combination of hyperparameters to be optimized is encoded as the position of the individual bat; through iterative optimization, the position of the individual bat that optimizes the fitness function is found, thereby determining the optimal combination of hyperparameters of the deep learning model.

8. The IoT time-series data prediction method based on intelligent swarm optimization deep learning according to claim 7, characterized in that, In step S4, the deep learning prediction model is specifically a long short-term memory network model or a variant thereof, including but not limited to gated recurrent unit networks or bidirectional long short-term memory networks; such models are suitable for prediction tasks of IoT time series data because they have the ability to capture long-term dependencies in time series data.

9. A IoT time-series data prediction system based on intelligent swarm optimization deep learning, characterized in that, The system employs the method described in any one of claims 1 to 8.

Images