Multi-modal streaming data prediction method and system

Abstract

The invention provides a multi-modal streaming data prediction method and system, and relates to the field of data prediction, and the method comprises the steps: describing the multi-modal distribution characteristics of streaming data according to a Gaussian mixture model; iterative updating is performed on Gaussian mixture model parameters based on an EM algorithm, modal identification is completed according to Gaussian mixture distribution, and an identification result is generated; according to each mode divided according to the identification result, establishing an independent SACRCJ prediction model and carrying out model training so as to send data belonging to the same mode into the corresponding SACRCJ model for prediction and obtain a local prediction result; and integrating all local prediction results based on a Bayesian method to obtain final prediction output. And the streaming data prediction efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of data prediction technology, and more specifically, to a multimodal streaming data prediction method and system. Background Technology

[0002] With the development of the Internet of Things (IoT) and intelligent systems, the scale of streaming data generation has surged. This type of data includes power load, environmental monitoring data, etc. Accurate prediction of such data is crucial for smart grid scheduling and resource optimization. Multimodal streaming data exhibits various local distribution characteristics due to changes in system operating modes, such as peak and off-peak periods and seasonal fluctuations in power load. Traditional single-modal prediction methods are no longer sufficient, necessitating modeling techniques adapted to multimodal characteristics. Current multimodal streaming data modeling faces two core challenges: first, how to accurately identify different modes of data to avoid local characteristics being masked by global averaging; and second, how to control computational complexity to ensure real-time performance while maintaining prediction accuracy. Although existing technologies have made breakthroughs in modality identification and prediction models, the lack of coordinated design between the two limits overall performance. The technical solutions most similar to this invention in existing technologies fall into three categories, as follows:

[0003] (1) Multimodal identification technology – Gaussian Mixture Model (GMM) application scheme. This scheme describes the distribution of multimodal data through a linear combination of multiple single Gaussian models. The core steps include: ① Initializing model parameters, specifically including the number of Gaussian components, mixing coefficients, mean, and covariance; ② Iteratively optimizing parameters using the expectation-maximization (EM) algorithm: E-step calculates the membership probability of a sample belonging to each Gaussian component, which is the mixing coefficient; M-step updates the mean, covariance, and mixing coefficient based on the membership probability until the log-likelihood function converges; ③ Assigning samples to the corresponding modes according to the maximum membership probability. This scheme can achieve probabilistic partitioning of multimodal data, but it can only complete identification and needs to be used in conjunction with an independent prediction model. The solution is to use existing known technologies, such as the application of GMM in the multi-condition monitoring method proposed by Ge Zhiqiang et al.

[0004] (2) Single-modal prediction technique – Reservoir computation (RC) model scheme. This scheme is illustrated using RC as an example. RC belongs to a simplified recurrent neural network, and its core structure includes an input layer, a reservoir, and an output layer: ① The input layer maps data to the reservoir through a weight matrix Win; ② The reservoir uses S-shaped neurons, updating the state through an internal weight matrix W and a feedback weight matrix Wback; ③ The output layer outputs the prediction result through an output weight matrix Wout, where Wout is solved using a ridge regression algorithm to avoid overfitting. This scheme has low computational complexity, but it is only suitable for single-modal data and cannot handle multimodal distribution differences. The solution is to improve existing RC techniques, such as the basic reservoir design of SACRCJ.

[0005] (3) Multimodal prediction technology – Gaussian process model (GPM) scheme. This scheme is based on Gaussian process theory to construct a multimodal prediction model. The core idea is: ① to regard multimodal data as a combination of multiple Gaussian processes and assign an independent kernel function to each mode; ② to infer the mean and uncertainty of the prediction output through Bayesian inference; ③ to directly model the global data without explicit mode identification. Although this scheme can characterize multimodal uncertainty, the computational complexity of the kernel function increases quadratically with the sample size, resulting in low prediction efficiency. The source is the publicly available GPM multimodal prediction method, which specifically shows that when there are 1000+ power load samples, the prediction time is more than 1.5 times that of this invention. In addition, there are also single-modal deep learning schemes such as LSTM in the existing technology, which fit the data pattern through deep networks, but the model training is time-consuming and has poor local adaptability to multimodal data, which easily leads to prediction deviations during peak and valley periods. The solution is the conventional application of existing deep learning in power load prediction.

[0006] In the process of developing this invention, the applicant discovered that the existing technology has at least the following shortcomings: In the field of multimodal streaming data prediction, existing modeling approaches have significant deficiencies. On the one hand, treating multimodal streaming data as a whole to establish a unified model, and only statistically averaging all modes, fails to accurately describe the local characteristics of the data, resulting in limited prediction accuracy. On the other hand, although some multimodal models, such as the Gaussian process model (GPM), can characterize the multimodality of data, these models have high computational complexity and low prediction efficiency. Furthermore, single-modal models such as SACRCJ, SCR, CESN, and LSTM cannot adapt to the distribution differences of multimodal data, easily leading to poor local fitting effects and large error fluctuations, especially in power load streaming data prediction, where it is difficult to balance accuracy and efficiency.

[0007] Therefore, improving the efficiency and accuracy of streaming data prediction has become a technical problem that needs to be solved. Summary of the Invention

[0008] This invention aims to solve at least one of the technical problems existing in the prior art or related technologies, and discloses a multimodal streaming data prediction method and system, which improves the efficiency and accuracy of streaming data prediction.

[0009] Explanation of technical terms:

[0010] Multimodal streaming data refers to data that is generated over time and exhibits multiple local statistical distributions due to changes in system operating modes (such as peak and valley loads and seasonal variations). Each distribution is called a "modality".

[0011] Reservoir Computing (RC): A simplified recurrent neural network with a fixed reservoir structure that requires no training; only the output weights are trained, thereby reducing model complexity.

[0012] SACRCJ: An improved RC model, characterized by simplifying the reservoir topology, introducing a cross-jump self-connected adjacent feedback loop topology, and using a hybrid neuron composed of S-type neurons and Symmlet wavelet neurons as the activation function to adapt to the real-time prediction requirements of streaming data.

[0013] EM algorithm: Expectation-Maximization algorithm, used to solve for the parameters of models such as GMM with latent variables, such as mode attribution, by iteratively approximating the optimal solution;

[0014] Bayesian integration: A result fusion method based on Bayesian probability theory, which integrates multiple local results through weight allocation to reduce the uncertainty of a single result.

[0015] The first aspect of this invention discloses a method for predicting multimodal streaming data, comprising: constructing a Gaussian mixture model: describing the multimodal distribution characteristics of streaming data according to the Gaussian mixture model, wherein the probability distribution ρ(x) of the Gaussian mixture model is represented by a linear combination of K Gaussian components:

[0016]

[0017] in, This represents the number of Gaussian components within a Gaussian mixture model. Indicates that the sample point comes from the th The mixing coefficient of the Gaussian components, Indicates the first The probability density function of Gaussian components;

[0018] Modality identification: Based on the EM algorithm, the parameters of the Gaussian mixture model are iteratively updated, and modality identification is completed according to the Gaussian mixture distribution to generate identification results;

[0019] Establish and train local SACRCJ prediction models: For each modality segmented based on the identification results, establish an independent SACRCJ prediction model and train it. This allows data belonging to the same modality to be fed into the corresponding SACRCJ model for prediction, obtaining local prediction results. The SACRCJ model is an improved RC model, employing a reservoir structure containing sigmoid neurons to reduce computational complexity. The SACRCJ model updates its state and outputs through the following process:

[0020]

[0021]

[0022] in, The input data at the current moment, and Represents the activation function. For the input weight matrix, and These represent the internal states of the reserve pool and their corresponding matrices. For the feedback weight matrix, The output data at the current moment. To output the weight matrix;

[0023] Integrating local prediction results: Based on the Bayesian method, all local prediction results are integrated to obtain the final prediction output.

[0024] The multimodal streaming data prediction method disclosed in this invention preferably further includes:

[0025] Data preprocessing: Streaming data arranged in chronological order is collected from real-time or historical databases. For these raw data with dimensional differences, a minimum-maximum normalization method is used to linearly transform them to the [0,1] interval. The streaming data includes power load data or environmental monitoring data. The normalization formula is as follows:

[0026]

[0027] in, This represents the minimum value of the input data to be predicted. To predict the maximum value of the input data, This is the output value of the data.

[0028] The multimodal streaming data prediction method disclosed in this invention preferably includes:

[0029] The parameters of the Gaussian mixture model are iteratively updated based on the EM algorithm. If the stopping condition of the EM algorithm is met, the modality identification is completed according to the Gaussian mixture distribution and the identification result is given. Otherwise, the EM algorithm iterative steps are repeated until the stopping condition of the EM algorithm is met.

[0030] Each iteration of the EM algorithm consists of two parts: an expectation-finding step and a maxima-finding step. These two parts are repeated continuously to obtain the final distribution parameters.

[0031] Expected solution steps: Based on the set initial parameter values Calculate samples Mixing coefficients from various Gaussian functions The calculation formula is as follows:

[0032]

[0033] in, Representing the The next iteration;

[0034] Maximization steps: Based on the mixing coefficients obtained in the expected solution steps The distribution parameters of the Gaussian mixture model are calculated using the following formula:

[0035]

[0036]

[0037]

[0038] in, , and They represent the first After the nth iteration The mean, covariance, and mixing coefficient of each Gaussian component. For the first Data points, Represents the total number of data points.

[0039] According to the multimodal streaming data prediction method disclosed in this invention, preferably, modality identification specifically includes:

[0040] Modality identification is a process that involves calculating and comparing samples. mixing coefficient To achieve this, according to Bayes' theorem, we can obtain:

[0041]

[0042] if , Then the sample He was assigned to the first In each modality, data identification under different distributions is completed at this point.

[0043] According to the multimodal streaming data prediction method disclosed in this invention, preferably, the step of integrating local prediction results specifically includes:

[0044] For a data point to be predicted, the posterior probability of the data point belonging to each mode is determined by the identification results, and the local predicted value of the data point is given by each local SACRCJ prediction model. Finally, the posterior probability is used as the weight to calculate the weighted average of all local predicted values ​​to obtain the final predicted value of the data point.

[0045] According to the multimodal streaming data prediction method disclosed in this invention, preferably, the local SACRCJ prediction model specifically includes:

[0046] The input layer receives input data and maps it to the high-dimensional space of the reserve pool using a randomly generated weight matrix.

[0047] The reservoir consists of multiple randomly connected and interacting neuron nodes, which update the internal state of the reservoir based on input data. The neuron nodes include S-shaped neuron nodes and wavelet neuron nodes.

[0048] The output layer receives all state information from the reservoir and maps the state information to the output target through a trainable linear weight matrix.

[0049] A second aspect of the present invention discloses a multimodal streaming data prediction system, comprising: a memory for storing program instructions; and a processor for calling the program instructions stored in the memory to implement the multimodal streaming data prediction method as described in any of the above technical solutions.

[0050] The beneficial effects of the present invention include at least the following:

[0051] 1. Compared with the existing solution of "GMM + independent single-modal model": This invention realizes deep collaboration between GMM and SACRCJ, so that the modality identification results of GMM can directly guide the parameter design of SACRCJ local model, avoiding the problem of the disconnect between modality identification and prediction in the existing solution, thereby significantly reducing prediction error, improving the fitting accuracy of local data characteristics, and better adapting to the distribution differences of multimodal data.

[0052] 2. Compared with the multimodal GPM scheme: This invention simplifies the reservoir structure design through SACRCJ, adopts a more efficient neuron connection and parameter solving method, significantly reduces computational complexity, has higher prediction efficiency under the same data scale, and can maintain good multimodal characterization capability, achieving a balance between "accuracy and efficiency".

[0053] 3. Compared with the single-modal SACRCJ / LSTM scheme: This invention effectively improves the fitting deviation of the single-modal model to the local characteristics of multimodal data by multimodal partitioning and local modeling, reduces the fluctuation of prediction error, and improves the overall prediction stability. The advantages are more prominent, especially in scenarios where there are obvious pattern changes in the data (such as during peak and valley periods).

[0054] 4. Enhanced adaptability in multi-step prediction: Existing models are prone to a sharp drop in accuracy during multi-step prediction. However, this invention, relying on modal adaptation and Bayesian result integration mechanisms, can effectively alleviate error accumulation in multi-step prediction and maintain high prediction reliability, meeting the needs of medium- and long-term prediction in practical applications. Attached Figure Description

[0055] Figure 1 A flowchart illustrating a multimodal streaming data prediction method according to an embodiment of the present invention is shown.

[0056] Figure 2 A schematic block diagram of a multimodal streaming data prediction system according to an embodiment of the present invention is shown.

[0057] Figure 3 The diagram shows the mode identification results of a power load streaming data of a power grid in Northwest China according to an embodiment of the present invention.

[0058] Figure 4 The prediction performance of GMM-SACRCJ in different modes is shown in an embodiment of the present invention.

[0059] Figure 5 The figure shows the prediction results of different models in the streaming data of a power load in the Northwest Power Grid according to an embodiment of the present invention.

[0060] Figure 6 The diagram illustrates the prediction error of different models in streaming data of a power load in the Northwest Power Grid according to an embodiment of the present invention.

[0061] Figure 7 A diagram illustrating the modal prediction process of streaming data for a power load in the Northwest Power Grid according to an embodiment of the present invention is shown.

[0062] Figure 8 The diagram illustrates the prediction results of different models in Australian electricity load streaming data according to an embodiment of the present invention.

[0063] Figure 9 A graph showing the prediction error of different models in Australian electricity load streaming data according to an embodiment of the present invention is illustrated.

[0064] Figure 10 A diagram illustrating the modal prediction process of Australian electricity load streaming data according to an embodiment of the present invention is shown.

[0065] Figure 11 A plot of submodal prediction error of Australian electricity load streaming data is shown according to an embodiment of the present invention.

[0066] Figure 12 The prediction performance of GMM-SACRCJ is shown in a graph based on an embodiment of the present invention, using Australian electricity load streaming data at different prediction step sizes.

[0067] Figure 13A schematic diagram of the topology of a SACRCJ model according to an embodiment of the present invention is shown.

[0068] Figure 14 A schematic diagram of the learning process of the SACRCJ model according to an embodiment of the present invention is shown. Detailed Implementation

[0069] To better understand the above-described objects, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention; however, the invention may be practiced in other ways different from those described herein, and therefore, the invention is not limited to the specific embodiments disclosed below.

[0070] An embodiment of the present invention discloses a multimodal streaming data prediction method, comprising:

[0071] Data preprocessing: Streaming data arranged in chronological order is collected from real-time or historical databases. For these raw data with dimensional differences, a minimum-maximum normalization method is used to linearly transform them to the [0,1] interval. The streaming data includes power load data or environmental monitoring data. The normalization formula is as follows:

[0072]

[0073] in, This represents the minimum value of the input data to be predicted. To predict the maximum value of the input data, The output value of the data;

[0074] Constructing a Gaussian Mixture Model (GMM): Based on the description of the multimodal distribution characteristics of streaming data using the Gaussian Mixture Model, the probability distribution ρ(x) of the Gaussian Mixture Model is represented by a linear combination of K Gaussian components:

[0075]

[0076] in, This represents the number of Gaussian components within a Gaussian mixture model. Indicates that the sample point comes from the th The mixing coefficient of the Gaussian components, Indicates the first The probability density function of Gaussian components;

[0077] Modality identification: Based on the EM algorithm, the parameters of the Gaussian mixture model are iteratively updated, and modality identification is completed according to the Gaussian mixture distribution to generate identification results;

[0078] Establish and train local SACRCJ prediction models: For each modality segmented based on the identification results, establish an independent SACRCJ prediction model and train it. This allows data belonging to the same modality to be fed into the corresponding SACRCJ model for prediction, obtaining local prediction results. The SACRCJ model employs a reservoir structure containing sigmoid neurons to reduce computational complexity. The SACRCJ model updates its state and outputs its results through the following process:

[0079]

[0080]

[0081] in, The input data at the current moment, and Represents the activation function. For the input weight matrix, and These represent the internal states of the reserve pool and their corresponding matrices. For the feedback weight matrix, The output data at the current moment. To output the weight matrix;

[0082] Integrating local prediction results: Based on the Bayesian method, all local prediction results are integrated to obtain the final prediction output.

[0083] like Figure 1 As shown, in this embodiment, considering the multimodal nature of complex streaming data in actual production and life, and the drawback of a single global model being easily affected by different data distributions, thus failing to accurately describe the global and local characteristics of streaming data, this embodiment provides a Gaussian Mixture RC Learning Model (GMM-SACRCJ) to solve this problem. GMM-SACRCJ first uses a Gaussian Mixture Model (GMM) to identify the modalities of the studied streaming data, then establishes a local prediction model based on SACRCJ, thereby achieving local prediction of multimodal streaming data. Finally, the local prediction results are integrated using a Bayesian method to obtain the final prediction result of the model. The streaming data prediction based on GMM-SACRCJ specifically includes the following steps:

[0084] Step 1, Data Preprocessing: Collect chronologically ordered streaming power load data from the target system's real-time or historical database. For these raw data with dimensional differences, use a min-max normalization method to linearly transform them to the [0,1] interval. Then, divide the data into training and testing data.

[0085] The normalization calculation formula is:

[0086]

[0087] in, This represents the minimum value of the input data to be predicted. To predict the maximum value of the input data, This is the output value of the data.

[0088] Step 2: Construct a Gaussian mixture model to describe the multimodal distribution characteristics of streaming data: The probability distribution of the GMM is represented by a linear combination of K Gaussian components, and its probability density function is:

[0089]

[0090] in, This represents the number of Gaussian components within a GMM. Indicates that the sample point comes from the th The mixing coefficients of Gaussian components, and satisfying . Indicates the first The probability density function of Gaussian components.

[0091] Step 3: Solve for parameters using the EM algorithm: Iterate and update the model parameters based on the EM algorithm. If the stopping condition of the EM algorithm is met, perform mode identification based on the Gaussian mixture distribution and give the identification result; otherwise, repeat the EM algorithm iteration steps. Each iteration of the EM algorithm is divided into two parts: the expectation solution step (E-step) and the maxima solution step (M-step), and these two parts are repeated continuously to obtain the final distribution parameters.

[0092] The specific steps of using the EM algorithm to solve the GMM are as follows:

[0093] E-step: Based on the set initial parameter values Calculate samples Mixing coefficients from various Gaussian functions The calculation formula is as follows:

[0094]

[0095] in, Representing the The next iteration.

[0096] M-step: The mixing coefficient obtained from the E-step. The distribution parameters of the GMM are calculated using the following formula:

[0097]

[0098]

[0099]

[0100] in, , and They represent the first After the nth iteration The mean, covariance, and mixing coefficient of each Gaussian component. For the first Data points, Represents the total number of data points.

[0101] Step 4, Modal segmentation of all streaming data: The trained GMM is used for modal segmentation of all streaming data. For each data point, the GMM is first used to identify the modality of the sample data and the sample data is divided into approximate modalities (cluster 1, cluster 2, ... cluster n).

[0102] The modality identification process mainly involves calculating and comparing samples. mixing coefficient To achieve this, according to Bayes' theorem, we can obtain:

[0103]

[0104] if , Then the sample He was assigned to the first In each modality, data identification under different distributions is completed at this point.

[0105] Step 5: For the different modalities of data, establish and train corresponding local SACRCJ prediction models: Establish an independent simplified adaptive reservoir computation joint model (SACRCJ model) for each modality. The SACRCJ model uses a reservoir structure containing only sigmoid neurons to reduce computational complexity. Data belonging to the same modality are fed into its dedicated SACRCJ model, and the model updates its state and outputs through the following process:

[0106]

[0107]

[0108] in, The input data at the current moment, and Represents the activation function. For the input weight matrix, and These represent the internal states of the reserve pool and their corresponding matrices. For the feedback weight matrix, The output data at the current moment. To output the weight matrix.

[0109] Step 6: Integrate all local prediction results (local output 1, local output 2, ..., local output n) based on the Bayesian method to obtain the final prediction output: For the data point to be predicted, the posterior probability of it belonging to each mode is determined by step 4, and its local prediction value is given by each local SACRCJ model in step 5. Finally, the posterior probability is used as the weight to perform a weighted average of all local prediction results to obtain the final prediction value of the data point.

[0110] like Figure 2 As shown, according to another embodiment of the present invention, a multimodal streaming data prediction system 200 is also disclosed, including: a memory 201 for storing program instructions; and a processor 202 for calling the program instructions stored in the memory to implement the multimodal streaming data prediction method as described in the above embodiments.

[0111] According to another embodiment of the present invention, the specific calculation process of the SACRCJ model is also disclosed, and the specific structure of SACRCJ is as follows: Figure 13 As shown. All nodes in its reservoir contain self-connections and are connected via feedback loop edges and bidirectional jump edges. Specifically, the jump step size is 2, meaning that odd-numbered nodes are bidirectionally jump-connected with adjacent odd-numbered nodes, and even-numbered nodes are bidirectionally jump-connected with adjacent even-numbered nodes.

[0112] The characteristic of the internal weight matrix of the model's reserve pool is: for any main diagonal elements child diagonal elements , , , Top right corner element bottom left element ,in, For example, when the reservoir size is 6, the weight matrix of all neuron connections within the reservoir is... It is expressed as follows:

[0113]

[0114] To further improve the model's prediction accuracy during SACRCJ construction, a method is proposed to replace some sigmoid neurons in the SACRCJ reservoir with Symmlet wavelet neurons, resulting in a reservoir with hybrid neurons. The reservoir state and model output update formulas are as follows:

[0115]

[0116]

[0117] in, For the internal state variables of the reserve pool, For the excitation function inside the reserve pool, It is the connection weight matrix between the input layer and the reservoir. It is the input vector of the model. This is the weight matrix for all neuron connections within the model. It is the feedback weight matrix between the output layer and the reserve pool. It is the model's output vector. The excitation function between the input and output units. This is the output weight matrix of the model, and also the only quantity that needs to be determined in the model. The mathematical expression for the selected Symmlet mother wavelet is as follows:

[0118]

[0119] Since each wavelet neuron uses a different transformation method, the wavelet function system generated by the Symmlet wavelet is as follows when the scaling factor and translation factor are different:

[0120]

[0121] in, Represents the scaling factor. Represents the translation factor. Represents the size of the reserve pool. This indicates the proportion of wavelet neurons in the mixed neurons.

[0122] Output weight matrix It can be obtained using the ridge regression algorithm, and its calculation formula is as follows:

[0123]

[0124] in, The row vectors of the output matrix. Let T be the row vector of the matrix formed by the internal state matrix and the input weight matrix of the reservoir, where T denotes the matrix transpose. Represents the regularization coefficient. It is an identity matrix.

[0125] The SACRCJ learning process mainly includes four parts: data processing, model building, model training, and model validation. The SACRCJ-based learning process is as follows: Figure 14 As shown.

[0126] (1) Data Processing. To better learn the patterns between data and adapt to the data output range of the model, normalization methods are usually used to process the data. After data processing, the data can be divided into two parts: training data and test data. At the same time, the training data and test data will be further divided into two parts: input data and target data.

[0127] (2) Model Construction. First, a simplified topological RC learning model, SACRCJ, is proposed. Since SACRCJ is constructed by simplifying and fixing the neuron connection method in the ESN model reservoir, this model will effectively reduce computational complexity and improve its regularity. Then, wavelet neurons are used to replace some of the S-type neurons in the SACRCJ reservoir to construct a hybrid neuron as the activation function of the model. The proposal of the hybrid neuron will effectively avoid the problem that singular solutions are easily generated when a single type of neuron is used as the activation function.

[0128] (3) Model Training. Model training begins by inputting the input data from the training data into SACRCJ to obtain the corresponding model training output. Then, the model's training output data is compared with the target data. If the error meets the iteration termination condition, the model training is complete, and a SACRCJ model with determined parameters is obtained. Otherwise, the data input step is repeated, and model training continues.

[0129] (4) Model validation. Model validation mainly involves inputting the input data from the test data into SACRCJ to obtain the corresponding test output, and comparing the test output data with the target data to obtain the corresponding test error.

[0130] According to another embodiment of the present invention, in order to verify the effectiveness and superiority of the multimodal streaming data prediction method provided in the above embodiments, sufficient experiments were conducted, and the experimental results are described in the following two parts:

[0131] (1) Effects of mode division and parameter determination:

[0132] This section aims to demonstrate that Gaussian mixture models (GMMs) can effectively identify different modes in streaming data and determine the optimal parameters for subsequent predictions.

[0133] Figure 3 This image shows the modal identification results of streaming data from a power load in the Northwest China Power Grid. The image illustrates the results after dividing the power load data of the Northwest China Power Grid into three modes (K=3). As shown in the figure, after GMM modal identification, the data points of different modes (distinguished by different colors) are neatly distributed, with clear boundaries between modes and low coupling. This visually demonstrates that the streaming data does indeed possess inherent multimodal characteristics and verifies the effectiveness of the first step, "modal identification," of this invention.

[0134] Figure 4 The graph shows the prediction performance of GMM-SACRCJ with different modalities. It illustrates the changing trends of the model's prediction performance metrics (NRMSE, RMSE, R²) as the number of modalities (K) increases from 1 to 4. Experimental results clearly show that when the number of modalities K=3, NRMSE and RMSE are minimized, while the R² value reaches its maximum of 0.9905. This graph provides a quantitative basis for determining the key parameters of this invention, proving that setting 3 modalities is the optimal choice for this dataset, thus ensuring the accuracy of subsequent local predictions.

[0135] Figure 7 This diagram illustrates the modal prediction process for streaming data of a power load in the Northwest Power Grid. It shows the prediction process under different modality settings: (a) C=1, the data is considered a single modality; (b) C=2, it is divided into two trends; (c) C=3, it is divided into three modes; (d) C=4, it is further subdivided. This diagram visually demonstrates the implementation process of the "divide and conquer" strategy, proving the necessity of multimodal partitioning to improve prediction accuracy.

[0136] Figure 10 A diagram illustrating the modal forecasting process of Australian electricity load streaming data, and... Figure 7 Correspondingly, the prediction results for different modality partitions on another dataset are shown, further verifying the universality of the method of the present invention.

[0137] Figure 11 The graph shows the prediction error of Australian electricity load streaming data by mode. The graph illustrates the change in prediction error when the number of modes is different. The results show that the prediction accuracy does not increase monotonically with the increase of the number of modes, but there is an optimal value (C=3). This provides an important basis for determining the optimal mode division.

[0138] (2) Comparison of overall model prediction performance:

[0139] This section compares the present invention (GMM-SACRCJ) with various advanced models, comprehensively demonstrating the superior performance of the present invention from three dimensions: prediction curve, error distribution, and quantification indicators.

[0140] Figure 5 The graph shows the prediction results for different models. Figure 5 As can be seen from the prediction curve, the prediction results of this invention (pink curve) fit the real data points (multi-colored points) most closely, especially at the turning points of data changes, where the fitting effect is significantly better than other single-modal models (SACRCJ, SCR, CESN, LSTM) and multi-modal comparison models (GPM).

[0141] Figure 6 This is a graph showing the prediction errors of different models. Figure 6 As can be seen from the prediction error, the prediction error curve of the present invention has the smallest fluctuation range, and the error values ​​are concentrated in the range of -0.1 to 0.1, which is much more stable than the error curves of other models. This proves that the prediction results provided by the present invention have higher accuracy and stability.

[0142] Figure 5 and Figure 6 These two sets of images are the core comparative evidence. Figure 8 and Figure 9 The two sets of graphs were repeated on another dataset for comparison. The results also showed that the prediction curve of the present invention is closest to the true value and has the smallest error fluctuation, further verifying the robustness and generalization ability of the present invention.

[0143] Figure 12 The graph shows the prediction performance of GMM-SACRCJ in Australian electricity load streaming data with different prediction step sizes. It illustrates the prediction effect as the prediction step size increases from 2 to 5 steps. While the prediction points become more dispersed with increasing step size, good trend-following ability is maintained even with 5-step prediction, demonstrating the good performance of this invention in multi-step prediction tasks.

[0144] Tables 1 and 2 present the prediction results of different models: these two tables provide the most crucial quantitative evidence. The data show that on two different datasets, the present invention (GMM-SACRCJ) comprehensively outperforms all comparative models in the core evaluation metrics NRMSE, RMSE, and R². For example, on the Northwest Power Grid data, the present invention achieves an R² of 0.9905, while other models only reach a maximum of 0.9715 (SACRCJ) or 0.9900 (GPM); on the Australian Power Grid data, the present invention achieves an R² of 0.98854. Experimental results demonstrate the significant advantage of the present invention in prediction accuracy.

[0145] Table 1. Prediction results of different models in the power load streaming data of a certain power grid in Northwest China.

[0146]

[0147] Table 2. Prediction results of different models in Australian electricity load streaming data.

[0148]

[0149] Table 3 shows the prediction results of GMM-SACRCJ for different prediction step lengths in Australian electricity load streaming data: This table provides the quantitative results of multi-step prediction, showing that even with 5-step prediction, the R² value can still reach 0.7559, proving that the model of the present invention maintains an acceptable level of accuracy in long-term prediction.

[0150] Table 3. Forecast results for different forecast step sizes in Australian electricity load streaming data.

[0151]

[0152] The above experimental results systematically demonstrate that the GMM-SACRCJ model proposed in this invention effectively solves the prediction problem of multimodal streaming data through the technical path of "modal identification → local prediction → result integration", and shows significant advantages in prediction accuracy, stability and generalization ability.

[0153] In summary, this invention presents a multimodal streaming data prediction architecture based on the collaboration of GMM and SACRCJ. GMM enables probabilistic modality identification, while SACRCJ adapts to each modality to construct local models, addressing the problems of "inaccurate global modeling and inefficient multimodal models" in existing technologies. This invention employs the parameter iterative optimization method based on the EM algorithm in GMM modality identification, combined with a dynamic modality allocation mechanism based on sample membership probabilities, ensuring the accuracy and adaptability of modality segmentation. Furthermore, this invention uses sigmoid neurons to reduce computational complexity and combines them with ridge regression to solve for output weights, achieving a balance between local prediction accuracy and efficiency.

[0154] All or part of the steps in the various methods of the above embodiments can be implemented by a program controlling the relevant hardware. The program can be stored in a readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other readable medium that can be used to carry or store data.

[0155] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multimodal streaming data prediction method, characterized in that, include: Constructing a Gaussian mixture model: The Gaussian mixture model describes the multimodal distribution characteristics of streaming data. The probability distribution ρ(x) of the Gaussian mixture model is represented by a linear combination of K Gaussian components. ； in, This represents the number of Gaussian components in a Gaussian mixture model. Indicates that the sample point comes from the th The mixing coefficient of the Gaussian components, Indicates the first The probability density function of Gaussian components; Modality identification: Based on the EM algorithm, the parameters of the Gaussian mixture model are iteratively updated, and modality identification is completed according to the Gaussian mixture distribution to generate identification results; Establish and train local SACRCJ prediction models: For each modality identified by the recognition results, establish an independent SACRCJ prediction model and train it to allow data belonging to the same modality to be fed into the corresponding SACRCJ model for prediction, thus obtaining local prediction results. The SACRCJ model is an improved RC model, employing a reservoir structure containing sigmoid neurons to reduce computational complexity. The SACRCJ model updates its state and outputs its results through the following process: ， ； in, The input data at the current moment, and Represents the activation function. For the input weight matrix, and These represent the internal states of the reserve pool and their corresponding matrices. For the feedback weight matrix, The output data at the current moment. To output the weight matrix; Integrating local prediction results: Based on the Bayesian method, all local prediction results are integrated to obtain the final prediction output.

2. The multimodal streaming data prediction method according to claim 1, characterized in that, Also includes: Data preprocessing: Streaming data arranged in chronological order is collected from real-time or historical databases. For these raw data with dimensional differences, a minimum-maximum normalization method is used to linearly transform them to the [0,1] interval. The streaming data includes power load data or environmental monitoring data. The normalization calculation formula is as follows: ； in, This represents the minimum value of the input data to be predicted. To predict the maximum value of the input data, This is the output value of the data.

3. The multimodal streaming data prediction method according to claim 1, characterized in that, The modality recognition steps specifically include: The parameters of the Gaussian mixture model are iteratively updated based on the EM algorithm. If the stopping condition of the EM algorithm is met, the modality identification is completed according to the Gaussian mixture distribution and the identification result is given. Otherwise, the EM algorithm iterative steps are repeated until the stopping condition of the EM algorithm is met. Each iteration of the EM algorithm consists of two parts: an expectation-finding step and a maxima-finding step. These two parts are repeated continuously to obtain the final distribution parameters. Expected solution steps: Based on the set initial parameter values Calculate samples Mixing coefficients from various Gaussian functions The calculation formula is as follows: ； in, Representing the The next iteration; Maximization steps: Based on the mixing coefficients obtained in the expected solution steps The distribution parameters of the Gaussian mixture model are calculated using the following formula: ， ， ； in, , and They represent the first After the nth iteration The mean, covariance, and mixing coefficient of each Gaussian component. For the first Data points, Represents the total number of data points.

4. The multimodal streaming data prediction method according to claim 3, characterized in that, The modality recognition specifically includes: Modality identification is a process that involves calculating and comparing samples. mixing coefficient To achieve this, according to Bayes' theorem, we can obtain: ； if , Then the sample He was assigned to the first In each modality, data identification under different distributions is completed at this point.

5. The multimodal streaming data prediction method according to claim 1, characterized in that, The step of integrating local prediction results specifically includes: For a data point to be predicted, the posterior probability of the data point belonging to each mode is determined by the identification result, and the local prediction value of the data point is given by each local SACRCJ prediction model. Finally, the posterior probability is used as the weight to calculate the weighted average of all local prediction values ​​to obtain the final prediction value of the data point.

6. The multimodal streaming data prediction method according to any one of claims 1 to 5, characterized in that, The local SACRCJ prediction model specifically includes: The input layer is used to receive input data and map the input data to the high-dimensional space of the reserve pool through a randomly generated weight matrix. The reservoir comprises multiple randomly connected and interacting neuron nodes, which update the internal state of the reservoir based on input data. The neuron nodes include S-shaped neuron nodes and wavelet neuron nodes. The output layer receives all state information from the reservoir and maps the state information to the output target through a trainable linear weight matrix.

7. A multimodal streaming data prediction system, characterized in that, include: Memory, used to store program instructions; A processor is configured to invoke the program instructions stored in the memory to implement the multimodal streaming data prediction method as described in any one of claims 1 to 6.

Images