Internet of Things perception data prediction method based on deep learning

Abstract

The invention relates to an Internet of Things perception data prediction method based on deep learning, and belongs to the technical field of Internet of Things. The method comprises the following steps: acquiring and preprocessing Internet of Things sensing long-term data, decomposing time sequence data by using an STL method according to the seasonal characteristic of the Internet of Things data, and extracting periodicity, tendency and residual error characteristics; the whale algorithm is improved from the perspective of a position updating strategy and prevention of falling into local optimum, and hyper-parameters of the bidirectional long-short-term memory network model are optimized by adopting the improved whale optimization algorithm; and utilizing the optimized BiLSTM model to carry out Internet of Things perception data prediction. The method can efficiently predict the future trend of the sensing data of the Internet of Things, is suitable for scenes such as flow monitoring and intelligent environment sensing, and provides reliable decision support for application of the Internet of Things.

Description

Technical Field

[0001] The present invention belongs to the technical field of Internet of Things and relates to a method for predicting Internet of Things perception data based on deep learning. Background Art

[0002] With the rapid development of the Internet of Things, a large amount of sensor data is generated and used in various application scenarios, such as smart homes, smart cities, industrial automation, and environmental monitoring. Timely and accurate prediction of these data is of great significance for achieving more efficient resource management, early warning, and improving the response speed of the system. However, since sensor data are usually nonlinear, non-stationary, and complex, traditional prediction methods are difficult to effectively process and predict complex sensor data patterns. In addition, sensor data are often affected by noise and missing data, which further limits the accuracy of data prediction.

[0003] In recent years, deep learning technology has shown significant advantages in processing large-scale and complex data. Through automatic feature extraction and hierarchical representation, deep learning algorithms can capture the deep structural features of data and are suitable for processing multi-dimensional and time-series data. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a deep learning-based IoT perception data prediction method, which has the advantages of high efficiency, accuracy and good robustness, so as to meet the above challenges and provide better prediction support for IoT applications.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] A method for predicting IoT perception data based on deep learning, the method comprising the following steps:

[0007] S1, obtain IoT sensing data and preprocess it using linear interpolation and CEEMDAN-Boxplot mutation point detection methods;

[0008] S2. Based on the seasonal characteristics of IoT data, STL is used to decompose the time series data and extract periodicity, trend and residual features;

[0009] S3, using nonlinear convergence factor, adaptive weight strategy and random difference mutation strategy to improve the whale algorithm;

[0010] S4. Establish a bidirectional long short-term memory network model and use the improved whale optimization algorithm to optimize hyperparameters;

[0011] S5. Take the periodicity, trend and residual features as input, predict the features through the bidirectional long short-term memory network model optimized by the improved whale optimization algorithm, and linearly add the prediction results to get the final prediction result.

[0012] Further, in step S1, linear interpolation is used to complete the collected IoT sensing data, wherein the linear interpolation method is expressed as:

[0013]

[0014] Among them, y k For missing data y k+n for y k The adjacent non-empty data after that, and y k Difference n data; y k-m for y k The previous adjacent non-empty data, and y k The difference is m data;

[0015] The CEEMDAN-Boxplot mutation point detection method is used to remove noise in the data. The process is as follows: the IoT perception data sample x(t) is decomposed by CEEMDAN to obtain an IMF signal and a residual component signal; the IMF signal frequency is distributed from high to low, and the IMF signal with an amplitude greater than the threshold in the high-frequency part of the IMF is removed by Boxplot, and then reconstructed to obtain the reconstructed data sample.

[0016]

[0017] Where M represents the total number of IMF signals decomposed by CEEMDAN, and the reconstructed data samples After removing the first k IMF signals after sorting, the value k is calculated as:

[0018]

[0019] Among them, DMSE(n) represents the discrete mean square error of the nth IMF component. High frequency and low frequency are distinguished by the discrete mean square error criterion, and its formula is:

[0020]

[0021] Where DMSE(n) represents the discrete mean square error of the nth IMF component; L is the number of original data; x(t) represents the tth data of the original signal; IMF n (t) represents the t-th data of the n-th IMF component; M is the number of IMF components.

[0022] Further, in step S2, the reconstructed data sample data As the time series data Y, seasonal trend decomposition is performed. Multiple iterations are achieved through nested double-layer loops. The decomposed trend and cycle are obtained through the inner loop. The outer loop updates a weight used to ensure the robustness of the algorithm. The process is expressed as follows:

[0023] Y=T t +S t +R t

[0024] Among them, Y represents the IoT perception data obtained after preprocessing, T t represents the trend item; S t Represents a periodic term; R t Represents the residual term.

[0025] Further, in step S3, the process of the improved whale optimization algorithm IWOA is:

[0026] S31. Initialize the whale population, including N randomly generated whale individuals, initialize the parameters, and set the maximum number of iterations t max , and set the initial round t = 1;

[0027] S32. In each round of iteration, a nonlinear convergence factor a is designed to update the values ​​of coefficient vectors A and C, and then the search method of the whale algorithm is controlled according to |A|;

[0028] S33, adopting an adaptive weight strategy to update different individual positions according to different individuals and different prey positions;

[0029] S34, update round t=t+1, and then update the individual position again through the random difference mutation strategy to obtain the optimal position before and after the change;

[0030] S35. Judgment: When t≥t max When t<t max , calculate the fitness of each individual, record the optimal individual position, and return to step S32.

[0031] Further, in step S32, the nonlinear convergence factor is expressed as:

[0032]

[0033] Among them, r 1 , r 2 are random numbers of [0, 1], and a is the improved nonlinear convergence factor;

[0034] The update formula of coefficient vectors A and C is expressed as;

[0035] A=2a·r 1 -a

[0036] C=2r 2

[0037] Among them, A is used to control the position update of the whale in the search space, and C is used to guide the attraction relationship between the whale and the current optimal solution. When |A|≤1, local search is adopted, and when |A|>1, global search is adopted.

[0038] Further, in step S33, first, a dynamic attenuation factor ω(t) is added, which gradually decreases as the number of iterations t increases, and its expression is:

[0039]

[0040] Among them, ω(t) is the nonlinear inertia weight, α(t) is the correction coefficient;

[0041] Randomly generate a value p in [0, 1]. If p<0.5, |A|<1, then:

[0042] X(t+1)=ω(t)·X p (t)-A·D

[0043] Among them, A is a coefficient vector, which is used to control the position update of the whale in the search space; D is a coefficient vector, which represents the distance between the current whale and the optimal solution; X p (t) is the prey position vector, X(t+1) is the individual position vector after updating at each stage;

[0044] If p<0.5, |A|≥1, then:

[0045] D=|C·X rand (t)-X(t)|,X(t+1)=X rand (t)-A·D

[0046] Among them, X rand is the position vector of a whale individual randomly selected from the current group;

[0047] If p ≥ 0.5, then:

[0048] X(t+1)=ω(t)·X p (t)+D·e bl ·cos(2πr)

[0049] Among them, e bl is a random constant, b is a constant used to define the shape of a logarithmic spiral, and l is a random number in the range of [-1, 1].

[0050] Further, in step S34, the process of updating the individual position again using the random difference mutation strategy is expressed as:

[0051] X(t+1)=l 1 ×(X p (t)-X(t))+l 2 ×(X * (txt))

[0052]

[0053] Among them, X * (t) is a random individual in the population, f best is the optimal fitness of the group, f avg is the average fitness of the group, l 1 and l 2 Represent the weight coefficients, l 1 Weight dominates local search, l 2 The weights dominate the global guide, ∈ is a minimum value that prevents division by zero.

[0054] Further, in step S4, the bidirectional long short-term memory network model includes a forward long short-term memory network layer and a backward long short-term memory network layer, wherein the input time series is X={x 1 , x 2 , ..., x t}, t is the length of the time series;

[0055] The forward LSTM is:

[0056]

[0057] in, is the hidden vector of the forward LSTM at time step t;

[0058] The backward LSTM is:

[0059]

[0060] in, is the hidden vector of the backward LSTM at time step t;

[0061] Final output:

[0062]

[0063] Among them, W forward is the weight matrix from the forward hidden state to the output, W backward is the weight matrix from the backward hidden state to the output, b y is the bias vector of the output layer, y tFinal output at time step t.

[0064] Furthermore, in step S4, the improved whale optimization algorithm IWOA is used to optimize the hyperparameters of the bidirectional long short-term memory network model, and the hyperparameters include: the number of hidden units H, the maximum training cycle E, and the learning rate η. The values ​​of H, E, and η are dynamically adjusted by IWOA, and the hyperparameters are optimized with the optimization goal of achieving the optimal balance between prediction accuracy and computational efficiency.

[0065] Further, in step S5, the optimized bidirectional long short-term memory network model converts the periodic term T t , Trend Item S t , residual term R t As input, and make corresponding predictions respectively, and finally add the three sets of prediction results linearly to get the final prediction result:

[0066]

[0067] in, Represents the results of the final combined model.

[0068] The beneficial effects of the present invention are:

[0069] 1. The present invention proposes a combined prediction method based on STL-IWOA-BiLSTM, which better captures the seasonal and trend characteristics of IoT perception data.

[0070] 2. Based on the WOA, the present invention improves the optimization effect of the algorithm by using nonlinear convergence factors, adaptive weight strategy and random difference method mutation strategy. IWOA has a higher convergence speed and global search capability, which enables the model to find more suitable data prediction parameters, avoids manual parameter adjustment, and improves the accuracy and stability of the model.

[0071] 3. The BiLSTM model that combines STL decomposition and IWOA optimization shows stronger robustness in dealing with complex, nonlinear and noisy data, and maintains high prediction stability in the case of missing data and abnormal conditions. This combined model can effectively deal with noise interference and missing data problems in different scenarios, thereby improving the reliability of predictions.

[0072] 4. The STL-IWOA-BiLSTM combined model of the present invention has good versatility and can adapt to the data prediction needs of various IoT scenarios, such as environmental monitoring, energy management, and intelligent manufacturing, and provide high-precision data prediction support for different applications.

[0073] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0075] Figure 1 A flow chart of the method provided by the present invention;

[0076] Figure 2 The figure is a flow chart of the improved whale algorithm IWOA of the present invention. DETAILED DESCRIPTION

[0077] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways 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 only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0078] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0079] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0080] See also Figure 1-2, which is a deep learning-based prediction method for IoT perception data.

[0081] Example

[0082] This embodiment proposes a specific implementation process of a method for predicting IoT perception data based on deep learning, such as Figure 1 As shown, it includes the following steps:

[0083] Step S1: Obtain IoT sensing data and preprocess it using linear interpolation and CEEMDAN-Boxplot mutation point detection methods;

[0084] Since IoT perception data is prone to data missing and data anomalies during the collection process, in order to better train the data, the linear interpolation method is used to complete the data and the CEEMDAN-Boxplot mutation point detection method is used to remove the influence of noise on the experimental results.

[0085] The linear interpolation method is as follows:

[0086]

[0087] Among them, y k For missing data y k+n for y k The adjacent non-empty data after that, and y k Difference n data; y k-m for y k The previous adjacent non-empty data, and y k There is a difference of m data.

[0088] The CEEMDAN-Boxplot mutation point detection method is as follows: perform CEEMDAN decomposition on the IoT perception data sample x(t) to obtain an IMF signal and a residual component signal. The IMF signal frequency is distributed from high to low. The high-frequency part of the IMF with large amplitude is removed through Boxplot, and then reconstructed. The high frequency and low frequency are distinguished by the discrete mean square error criterion, and the formula is:

[0089]

[0090] Where: DMSE(n) represents the discrete mean square error of the nth IMF component; L is the number of original data; x(t) represents the tth data of the original signal; IMF n (t) represents the t-th data of the n-th IMF component; M is the number of IMF components.

[0091] DMSE can represent the reconstruction error between the IMF component and the original signal. The IMF component with a smaller DMSE value corresponds to the main component of the original signal, while the IMF component with a larger DMSE value corresponds to noise or subtle details. A value k is calculated by the following formula, so that the part after the kth IMF component is a low-frequency signal and retains valuable signal information, while the high-frequency component mainly contains noise and mutation point signals. The expression is:

[0092]

[0093] Reconstructed data samples after removing noise and mutation point signals It is expressed as:

[0094]

[0095] Step S2: Use seasonal and trend decomposition using Loess (STL) to decompose the preprocessed time series data To decompose;

[0096] IoT sensing data usually has seasonal and trend characteristics. Using STL decomposition, the data can be decomposed into periodic terms, trend terms, and residual terms. The periodic component is relatively stable and accounts for a large proportion of the data, while the trend component is unstable. Some data have obvious trends, while others do not. Therefore, directly training and predicting the initial time series data will weaken the trend component during the model learning process, and the final prediction result may deviate from the original trend.

[0097] The STL decomposition formula is as follows:

[0098] Y=T t +S t +R t

[0099] Among them, Y represents the IoT perception data obtained after preprocessing, T t represents the trend term; S t Represents a periodic term; R t Represents the residual term. The decomposition is implemented through multiple iterations through nested double loops. The decomposed trend and cycle can be obtained through the inner loop, while the outer loop updates a weight to ensure the robustness of the algorithm.

[0100] Step S3: Improve the Improved Whale Optimization Algorithm (IWOA) by using nonlinear convergence factor, adaptive weight strategy and random difference mutation strategy respectively; Figure 2The improved whale algorithm IWOA flow chart of the present invention is shown, which includes the following steps:

[0101] Step S3.1: Form a whale group, including N randomly generated whale individuals, initialize the parameters, and set the maximum number of iterations t max , let t = 1;

[0102] Step S3.2: Update the values ​​of A and C according to the proposed nonlinear convergence factor. A and C are coefficient vectors. A is used to control the position update of the whale in the search space, and C is used to guide the attraction relationship between the whale and the current optimal solution. In WOA, whether to use global random search or local search is determined according to the value of |A|. In the early stage of iteration, more global search is required, and in the later stage, more local search is required. In the original WOA, the size of |A| is controlled by the a value that decreases linearly from 2 to 0, which cannot control the global and local searches well. Therefore, the present invention proposes a nonlinear convergence factor that decreases nonlinearly from 2 to 0, as shown in the formula:

[0103]

[0104] A=2a·r 1 -a

[0105] C=2r 2

[0106] Among them, r 1 , r 2 are random numbers of [0, 1], a is the improved nonlinear convergence factor, and the size of |A| is controlled according to the value of a; when |A|≤1, local search is used, and when |A|>1, global search is used. The coefficient a converges slowly in the early stage, which can expand the search range and improve the global search ability of the algorithm; the convergence speed is fast in the later stage, which can improve the solution efficiency, make the local mining ability of the algorithm stronger, and improve the search ability of the algorithm while accelerating the convergence speed.

[0107] Step S3.3: Randomly generate a value p in [0, 1], and update it differently according to different individuals and different prey positions, using an adaptive weight strategy to update the position. Add a dynamic attenuation factor ω(t), which gradually decreases as the number of iterations t increases. Its expression is:

[0108]

[0109] Where ω(t) is the nonlinear inertia weight, t max is the maximum number of iterations, and α(t) is the correction coefficient.

[0110] The spiral update phase expression is then:

[0111] When p<0.5, |A|<1:

[0112] X(t+1)=ω(t)·X p (t)-A·D

[0113] Among them, A is a coefficient vector used to control the position update of the whale in the search space, D is a coefficient vector representing the distance between the current whale and the optimal solution, and X p (t) is the prey position vector, and X(t+1) is the individual position vector after updating at each stage.

[0114] When p<0.5, |A|≥1:

[0115] D=|C·X rand (t)-X(t)|,X(t+1)=X rand (t)-A·D

[0116] Where: X rand is the position vector of a whale randomly selected from the current group.

[0117] When p ≥ 0.5:

[0118] X(t+1)=ω(t)·X p (t)+D·e bl ·cos(2πr)

[0119] Among them, e bl is a random constant, b is a constant used to define the shape of a logarithmic spiral, and l is a random number in the range of [-1, 1].

[0120] Step S3.4: The individual updates it again through the random difference mutation strategy, taking the optimal position before and after the change, and setting t = t + 1, so that the population convergence is accelerated, effectively preventing the population from falling into the local optimum. Its expression is:

[0121] X(t+1)=l 1 ×(X p (t)-X(t))+l 2 ×(X * (txt))

[0122]

[0123] Among them, X * (t) is a random individual in the population, f best is the optimal fitness of the group, f avg is the average fitness of the group, l 1 and l 2 Represent the weight coefficients, l 1Weight dominates local search, l 2 The weights dominate the global guide, ∈ is a minimum value that prevents division by zero.

[0124] Step S3.5: Determine when t ≥ t max When t<t max , calculate the fitness of each individual and record the optimal individual position.

[0125] Step S4: Establish a bidirectional long short-term memory network (BiLSTM) model, which is used to predict IoT perception data based on the extracted periodic term, trend term and residual term features.

[0126] The bidirectional long short-term memory network model includes a forward long short-term memory network layer and a backward long short-term memory network layer, where:

[0127] The input time series is X = {x 1 , x 2 , ..., x t}, t is the length of the time series.

[0128] The forward LSTM is:

[0129]

[0130] in, is the hidden vector of the forward LSTM at time step t.

[0131] The backward LSTM is:

[0132]

[0133] in, is the hidden vector of the backward LSTM at time step t.

[0134] Final output:

[0135]

[0136] Where: W forward is the weight matrix from the forward hidden state to the output, W backward is the weight matrix from the backward hidden state to the output, b y is the bias vector of the output layer, y t Final output at time step t.

[0137] The IWOA algorithm in step S3 is used to optimize BiLSTM. The IWOA algorithm is used to optimize the number of hidden units, maximum training cycle, learning rate and other parameters in BiLSTM through grid search.

[0138] S5: The BiLSTM model optimized by the IWOA algorithm takes the periodic term, trend term, and residual term obtained by STL decomposition in step S2 as input, and makes corresponding predictions respectively. Finally, the three sets of prediction results are linearly added to obtain the final prediction result.

[0139]

[0140] in, Represents the results of the final combined model.

[0141] This embodiment also provides the following evaluation indicators for the final STL-IWOA-BiLSTM combined prediction model:

[0142]

[0143]

[0144] Among them, yi is the true value of the predicted object, is the predicted value after denormalization, is the actual mean of the predicted object, and n is the amount of data.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for predicting IoT perception data based on deep learning, characterized in that: The method comprises the following steps: S1, obtain IoT sensing data and preprocess it using linear interpolation and CEEMDAN-Boxplot mutation point detection methods; S2. Based on the seasonal characteristics of IoT data, STL is used to decompose the time series data and extract periodicity, trend and residual features; S3, using nonlinear convergence factor, adaptive weight strategy and random difference mutation strategy to improve the whale algorithm; S4. Establish a bidirectional long short-term memory network model and use the improved whale optimization algorithm to optimize hyperparameters; S5. Take the periodicity, trend and residual features as input, predict the features through the bidirectional long short-term memory network model optimized by the improved whale optimization algorithm, and linearly add the prediction results to get the final prediction result.

2. The method for predicting IoT perception data based on deep learning according to claim 1, characterized in that: In step S1, linear interpolation is used to complete the collected IoT sensing data, where the linear interpolation method is expressed as: Among them, y k For missing data y k+n for y k The adjacent non-empty data after that, and y k Difference n data; y k-m for y k The previous adjacent non-empty data, and y k There are m data differences; The CEEMDAN-Boxplot mutation point detection method is used to remove noise in the data. The process is as follows: the IoT perception data sample x(t) is decomposed by CEEMDAN to obtain an IMF signal and a residual component signal; the IMF signal frequency is distributed from high to low, and the IMF signal with an amplitude greater than the threshold in the high-frequency part of the IMF is removed by Boxplot, and then reconstructed to obtain the reconstructed data sample. Where M represents the total number of IMF signals decomposed by CEEMDAN, and the reconstructed data samples After removing the first k IMF signals after sorting, the value k is calculated as: Among them, DMSE(n) represents the discrete mean square error of the nth IMF component. High frequency and low frequency are distinguished by the discrete mean square error criterion, and its formula is: Where DMSE(n) represents the discrete mean square error of the nth IMF component; L is the number of original data; x(t) represents the tth data of the original signal; IMF n (t) represents the t-th data of the n-th IMF component; M is the number of IMF components.

3. The method for predicting IoT perception data based on deep learning according to claim 2, characterized in that: In step S2, the reconstructed data sample data As the time series data Y, seasonal trend decomposition is performed. Multiple iterations are achieved through nested double-layer loops. The decomposed trend and cycle are obtained through the inner loop. The outer loop updates a weight used to ensure the robustness of the algorithm. The process is expressed as follows: Y=T t +S t +R t Among them, Y represents the IoT perception data obtained after preprocessing, T t represents the trend item; S t Represents a periodic term; R t Represents the residual term.

4. The method for predicting IoT perception data based on deep learning according to claim 3 is characterized in that: In step S3, the process of the improved whale optimization algorithm IWOA is: S31. Initialize the whale population, including N randomly generated whale individuals, initialize the parameters, and set the maximum number of iterations t max , and set the initial round t = 1; S32. In each round of iteration, a nonlinear convergence factor a is designed to update the values ​​of coefficient vectors A and C, and then the search method of the whale algorithm is controlled according to |A|; S33, adopting an adaptive weight strategy to update different individual positions according to different individuals and different prey positions; S34, update round t=t+1, and then update the individual position again through the random difference mutation strategy to obtain the optimal position before and after the change; S35. Judgment: When t≥t max When t <t max , calculate the fitness of each individual, record the optimal individual position, and return to step S32.

5. The method for predicting IoT perception data based on deep learning according to claim 4 is characterized in that: In step S32, the nonlinear convergence factor is expressed as: Among them, r1, r2 are random numbers in [0,1], and a is the improved nonlinear convergence factor; The update formula of coefficient vectors A and C is expressed as; A=2a·r1-a C＝2r2 Among them, A is used to control the position update of the whale in the search space, and C is used to guide the attraction relationship between the whale and the current optimal solution. When |A|≤1, local search is adopted, and when |A|>1, global search is adopted.

6. The method for predicting IoT perception data based on deep learning according to claim 5, characterized in that: In step S33, first, a dynamic attenuation factor ω(t) is added, which gradually decreases as the number of iterations t increases. Its expression is: Among them, ω(t) is the nonlinear inertia weight, α(t) is the correction coefficient; Randomly generate a value p in [0,1]. If p<0.5, |A|<1, then: X(t+1)=ω(t)·X p (t)-A·D Among them, A is a coefficient vector, which is used to control the position update of the whale in the search space; D is a coefficient vector, which represents the distance between the current whale and the optimal solution; X p (t) is the prey position vector, X(t+1) is the individual position vector after updating at each stage; If p<0.5, |A|≥1, then: D=∣C·X rand (t)-X(t)∣,X(t+1)=X rand (t)-A·D Among them, X rand is the position vector of a whale individual randomly selected from the current group; If p ≥ 0.5, then: X(t+1)=ω(t)·X p (t)+D·e bl ·cos(2πr) Among them, e bl is a random constant, b is a constant used to define the shape of the logarithmic spiral, and l is a random number in the range of [-1,1].

7. The method for predicting IoT perception data based on deep learning according to claim 6, characterized in that: In step S34, the process of updating the individual position again using the random difference mutation strategy is expressed as: X(t+1)=l1×(X p (t)-X(t))+l2×(X * (t)-X(t)) Among them, X * (t) is a random individual in the population, f best is the optimal fitness of the group, f avg is the average fitness of the group, l1 and l2 represent weight coefficients respectively, l1 weight dominates local search, l2 weight dominates global guidance, ∈ is the minimum value to prevent division by zero.

8. The method for predicting IoT perception data based on deep learning according to claim 4, characterized in that: In step S4, the bidirectional long short-term memory network model includes a forward long short-term memory network layer and a backward long short-term memory network layer, wherein the input time series is X = {x1, x2, ..., x t }, t is the length of the time series; The forward LSTM is: in, is the hidden vector of the forward LSTM at time step t; The backward LSTM is: in, is the hidden vector of the backward LSTM at time step t; Final output: Among them, W forward is the weight matrix from the forward hidden state to the output, W backward is the weight matrix from the backward hidden state to the output, b y is the bias vector of the output layer, y t Final output at time step t.

9. The method for predicting IoT perception data based on deep learning according to claim 8, characterized in that: In step S4, the improved whale optimization algorithm IWOA is used to optimize the hyperparameters of the bidirectional long short-term memory network model, including: the number of hidden units H, the maximum training cycle E and the learning rate η; the values ​​of H, E and η are dynamically adjusted by IWOA, and the hyperparameters are optimized with the optimization goal of achieving the optimal balance between prediction accuracy and computational efficiency.

10. The method for predicting IoT perception data based on deep learning according to claim 9, characterized in that: In step S5, the optimized bidirectional long short-term memory network model converts the periodic term T t , Trend Item S t , residual term R t As input, and make corresponding predictions respectively, and finally add the three sets of prediction results linearly to get the final prediction result: in, Represents the results of the final combined model.

Images