Method for channel water level prediction based on online kernel echo state network of unsupervised learning filter

Abstract

The application provides a channel water level prediction method based on an online kernel echo state network of an unsupervised learning filter, relates to the technical field of water level prediction, and solves the problem of model prediction uncertainty by designing a kernel reserve state through a kernel method to replace traditional random weight initialization; an online learning mechanism of dynamically increasing hidden neurons according to new input data is further provided, the limitation of a fixed traditional neural network structure is broken through, the model can adaptively learn new water level change rules, a non-supervised learning filter based on Mean-shift and KD-tree is introduced, the distance between new data and an existing knowledge center point is calculated in real time, and it is intelligently decided whether new neurons are increased to learn new features, only connection weights are updated to fine-tune the model, or redundant data is deleted, so that the model is automatically and online updated and constructed, unlimited expansion of the network structure is avoided, and the precision and efficiency of medium and long term water level prediction are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of water level prediction technology, and in particular to a method for predicting water levels in waterways based on an online kernel echo state network using an unsupervised learning filter. Background Technology

[0002] Waterway transportation, as a crucial mode of transport, relies heavily on accurate water level forecasting for its safety and efficiency. With the continuous advancement of waterway information technology, water level observation stations deployed along waterways can collect massive amounts of time-series data in real time, providing a rich data foundation for water level forecasting. In recent years, artificial intelligence technology, especially deep neural network models, has demonstrated significant potential in the field of water level forecasting, effectively uncovering nonlinear relationships and time dependencies in the data to achieve higher-precision predictions.

[0003] However, current AI-based water level prediction methods, especially mainstream deep neural network models, have revealed significant shortcomings: firstly, the uncertainty of model predictions; and secondly, the lack of an efficient online self-updating mechanism. Particularly problematic is that once these models complete initial training, their network architecture (including the number of layers, neurons, and connection patterns) becomes fixed. Subsequent "updates" and "learning" are merely limited to minor adjustments to the connection weights between neurons within the established architecture. This rigid architecture and limited update mechanism makes it difficult for the model to adapt to dynamic water level changes based on real-time new data. For example, the inventor's prior patent application CN118586557A – A Method for Predicting Channel Water Levels Based on an Online Echo State Network Algorithm – proposes an online sequential echo state network model. This technology selects input weights and internal weights through a self-attention mechanism and scaling method, and utilizes an online learning network to update the model. Although this method solves the problem of model retraining to some extent, the randomness of its initial weights still leads to prediction uncertainty, and the update is limited to weight adjustment, which cannot adapt to the structural changes in water level data and cannot guarantee that the updated model can learn new water level change characteristics and patterns. Summary of the Invention

[0004] In view of this, this invention provides a channel water level prediction method based on an online kernel echo state network using an unsupervised learning filter. Utilizing the time-series characteristics of water level data, a water level prediction model structure is constructed. Based on newly generated water level data, the hidden neurons in the neural network are selected for generation or updating, enabling the autonomous construction and training of a novel water level prediction model. This overcomes the limitations of fixed model structures and updating only weights during updates, solving the problem of prediction uncertainty in model prediction and achieving high-precision prediction of medium- and long-term water levels. It can not only predict water levels using historical information but also autonomously construct and update the model using new data to adapt to new water level change patterns.

[0005] Therefore, the present invention provides the following technical solution:

[0006] This invention provides a method for predicting channel water levels based on an online kernel echo state network using unsupervised learning filters. The method includes:

[0007] S1. Obtain water level data of the waterway recorded by multiple water level observation stations deployed along the waterway;

[0008] S2. Construct and train an online kernel echo state network based on an unsupervised learning filter as a medium- to long-term channel water level prediction model; during model training and updating, find the center point of the training reserve memory according to the unsupervised learning method, compare the center point with the new data, determine the relationship between the new data and the reserve memory features, and determine whether to add new neurons to learn new features, update only the connection weights to fine-tune the model, or delete redundant data based on the judgment result; the kernel method is used to calculate the reserve pool state in the echo state network;

[0009] S3. Based on the acquired water level data, use the trained water level prediction model to predict the water level of the waterway.

[0010] Furthermore, the center point of the training reserve memory is found based on unsupervised learning methods, including finding the center point of the training reserve memory through Mean-shift clustering and KD-tree spatial indexing techniques.

[0011] Furthermore, the center points of the training reserve memory are found using Mean-shift clustering and KD-tree spatial indexing techniques, including:

[0012] The distance between a new input feature and existing features is obtained by calculating the minimum Euclidean distance between the new input feature and each cluster center, expressed as:

[0013] ;

[0014] in, Indicates the first i From the initial training features to the...n The distance between the center points; It is the first n One center point ( n = [1,2, , n ]); For the first i One memory feature; It is the Euclidean distance function;

[0015] For the current set of all memory features Construct a KD-Tree and recursively partition the data space;

[0016] For each candidate center point Use KD-Tree for fast range queries;

[0017] In the k In the next iteration, the search is performed to find the point that falls within the current center point. Centered on, with radius All neighboring points within the hypersphere are represented as:

[0018] ;

[0019] in, Based on the n The first center point k An index containing samples with the most similar features;

[0020] The new center point location, i.e., the mean of all points in that neighborhood, is calculated as follows:

[0021] ;

[0022] in, Indicates the updated center point; Indicates the first in the neighborhood i The process continues until the center point converges or the maximum number of iterations is reached; the final set of stable center points is the density pattern center of the memory feature.

[0023] Furthermore, by comparing the center point with the newly added data, the relationship between the newly added data and the reserved memory features is determined, and the model weights are updated based on the determination results, including:

[0024] When there are new input features Upon entry, calculate the new input features and the features based on the current memory. mem Calculated center point Distance:

[0025] ;

[0026] if If so, then delete the new input feature. No model updates are performed;

[0027] if Then, for this new input feature The model update process only adjusts the connection weights of the neural network without increasing the number of neurons; among them, It is the set distance threshold;

[0028] if This increases the number of neurons in the neural network.

[0029] Furthermore, the number of neurons in the neural network is increased, and the connection weights of the neural network are adjusted, including:

[0030] The first nuclear stockpile status Calculate using the following formula:

[0031] ;

[0032] in, This represents the first input feature in the training data; ( ) is the Gaussian kernel function; This indicates the memory feature selected by the model; in this case, the memory feature is the first input feature.

[0033] Based on intermediate variables Initialized output weights Use the following formula to perform the calculation:

[0034] ;

[0035] in, The target value corresponding to the first trained feature. To initialize intermediate variables:

[0036] ;

[0037] Based on the newly entered input features Nuclear stockpile status Calculation:

[0038] ;

[0039] in, It is the spectral radius coefficient. These are the memory features selected by the model, expressed as ;

[0040] For intermediate variables Update:

[0041] ;

[0042] in, ;

[0043] Calculate the error using the output weights from the previous step:

[0044] ;

[0045] The current output weight is calculated based on the error. Update:

[0046] ;

[0047] Through the above update process, whenever a new input feature enters the model, the predictive neural network will automatically add corresponding neurons according to the input feature, enabling the model to learn new predictive characteristics and patterns.

[0048] Furthermore, when a new input feature enters the training process, the information transfer variable is calculated. The calculation formula is:

[0049] ;

[0050] in, ;

[0051] Based on the information transmission variables, the output weights at this stage are updated according to the following formula:

[0052] ;

[0053] in, For the first i The actual value of the target corresponding to each step.

[0054] Furthermore, the channel level prediction model is used to predict the channel level, including: predicting the channel level using the updated output weights, as shown below:

[0055] ;

[0056] in, For the first The predicted water level of the waterway corresponding to each step; These are the updated output weights obtained after model training. It is an intermediate variable.

[0057] Advantages and positive effects of the present invention:

[0058] 1) This invention proposes a novel online water level prediction method. Based on the generated new water level data, the method selects to generate or update the hidden neurons in the neural network, thereby realizing the autonomous construction and training of the water level prediction model to adapt to the new water level change patterns.

[0059] 2) This invention utilizes the proposed unsupervised learning feature filtering method to guide model updates, thereby reducing the number of similar features entering the model knowledge base, improving training efficiency, and avoiding the problem of excessive computation caused by model size expansion.

[0060] 3) This invention utilizes the proposed nuclear reserve pool state representation method to effectively reduce the uncertainty of prediction results and enhance the reliability of the model.

[0061] 4) The water level prediction model of the present invention can be adapted to various scales of data for model training and deployment of service interfaces, providing convenient water level data for waterway bureaus or crew members, improving the level of waterway information services and ship navigation safety. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart of a channel water level prediction method based on an online kernel echo state network with an unsupervised learning filter, as described in an embodiment of the present invention.

[0064] Figure 2 This is a comparison of the predicted and actual values ​​of the OSK-ESN-MFKD model in the seventh step of the artificial data in this embodiment of the invention.

[0065] Figure 3 This is a comparison of the performance of OSK-ESN-MFKD with other online prediction models in predicting artificial data in this embodiment of the invention. Detailed Implementation

[0066] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0067] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0068] This invention proposes an Online Sequential Kernel Echo State Network (OSK-ESN-MFKD) with a Mean-shift KD-tree filter for medium- to long-term multi-state water level prediction. First, this technique designs a novel kernel reserve state using a kernel method to replace traditional random weight initialization, effectively solving the problem of model prediction uncertainty. Second, it proposes an online learning mechanism that dynamically adds hidden neurons based on new input data, breaking through the limitations of fixed traditional neural network structures and enabling the model to adaptively learn new water level change patterns. Furthermore, it innovatively introduces an unsupervised learning filter based on Mean-shift and KD-tree. By calculating the distance between new data and existing knowledge centers in real time, it intelligently decides whether to add new neurons to learn new features, update connection weights only to fine-tune the model, or delete redundant data. This achieves autonomous online model updates and construction while avoiding the infinite expansion of the network structure, significantly improving the accuracy and efficiency of medium- to long-term water level prediction.

[0069] like Figure 1 As shown, this embodiment of the invention provides a method for predicting channel water levels based on an online kernel echo state network using an unsupervised learning filter. The method specifically includes the following steps:

[0070] S1. Obtain water level data of the waterway recorded by multiple water level observation stations deployed along the waterway;

[0071] S2. Construct and train an online kernel echo state network based on unsupervised learning filters as a medium- to long-term channel water level prediction model;

[0072] In practical implementation, during model training and updating, the model can find the center point of the training reserve memory using unsupervised learning methods. By comparing the center point with newly added data, the relationship between the new data and the reserve memory features is determined. If the features of the new data are dissimilar to those of the reserve memory, new neurons are added to the prediction neural network; otherwise, an update strategy is used to update the connection weights of the model. Secondly, in constructing the neural network, the kernel method is used to improve the calculation of the reserve pool state, replacing the random selection method of weights to address the problem of prediction uncertainty. Its prediction results and self-updating characteristics can closely follow the water level change patterns, thus providing reliable data support for waterway safety management and flood and drought early warning.

[0073] The core of this method lies in replacing the random weight selection mechanism in traditional reservoir computing with a kernel method, fundamentally solving the uncertainty problem in model training and prediction. Simultaneously, it introduces an online data filtering mechanism based on unsupervised learning, dynamically evaluating the correlation between new and historical data through Mean-shift clustering and KD-tree spatial indexing techniques, thereby intelligently deciding on the addition or removal of neurons or the updating of weights. This innovative architecture not only retains the efficient computational characteristics of traditional echo state networks but also, through the deterministic mapping of the kernel method and the dynamic adjustment capability of unsupervised learning, enables the model to autonomously adapt to the complex spatiotemporal characteristics of water level changes, significantly improving the accuracy and stability of medium- and long-term water level prediction. It leverages the advantages of reservoir computing (RC), addressing the instability problem of prediction using the kernel method in addition to improving the accuracy of medium- and long-term water level prediction, and implementing a model structure that can automatically adjust the neural network results based on new data. This solves the problem that neural network results cannot be autonomously constructed and updated, making it better applicable to real-world scenarios.

[0074] S3. Based on the acquired water level data, use the waterway water level prediction model to predict the waterway water level.

[0075] In OSK-ESN-MFKD, the model is built using training data during the training process. First, the state of the first kernel reserve pool... The following formula can be used for calculation:

[0076] (1)

[0077] in, This represents the first input feature in the training data; ( ) is the Gaussian kernel function; This indicates the memory feature selected by the model; in this case, the memory feature is the first input feature.

[0078] Based on intermediate variables Initialized output weights The calculation can be performed using formula (2):

[0079] (2)

[0080] in, The target value corresponding to the first trained feature. To initialize intermediate variables:

[0081] .

[0082] Next, this invention designs a method for increasing the number of neurons in a neural network based on new input features. Nuclear stockpile status The calculation process can be expressed by formula (3):

[0083] (3)

[0084] in, It is the spectral radius coefficient (generally taking values ​​between [0, 1]). These are the memory features selected by the model, which can be expressed as .

[0085] To construct new neurons in a neural network from new input features, thereby increasing the number of neurons and enabling these new neurons to express new water level features, the intermediate variables... The update process can be expressed by formula (4).

[0086] (4)

[0087] in, In this step, in order to eliminate the error of the previous step, the output weight of the previous step is used to calculate the error, which can be calculated using formula (5).

[0088] (5)

[0089] Then, the current output weights are calculated based on the error. The update of can be expressed mathematically by formula (6).

[0090] (6)

[0091] Through the above update process, whenever new input features enter the model, the predictive neural network automatically adds corresponding neurons based on the input features, enabling the model to learn new predictive characteristics and patterns. However, if the training data is too large, using this method to update the model will cause the number of neurons in the predictive model to grow indefinitely, and the subsequent computational load will increase exponentially.

[0092] To address this issue, this invention proposes a feature filtering method based on unsupervised learning. This method utilizes unsupervised learning to identify the clustering structure of learned, memorized features, and then compares the clustering with new input features. If the distance D exceeds a set threshold π, it indicates that the new feature is significantly different from the existing memorized features, and the model can be updated by adding neurons to learn the new feature and its changing patterns. If the distance satisfies 0.2 ≤ D < π, the new feature is incorporated into the model by adjusting the neural network connection weights. If D < 0.2, the new feature is considered to be duplicated with existing features, and to avoid overfitting, this feature is removed and not included in model training. The specific method and update process are as follows:

[0093] First, to address the computationally intensive and slow iteration speed of the traditional Mean-Shift algorithm, this invention employs an acceleration method based on the spatial index structure KD-tree to efficiently identify density pattern centers from memory features. The distance between a new input feature and existing features can be obtained by calculating its minimum Euclidean distance to each cluster center, expressed as:

[0094] (7)

[0095] in, Indicates the first i From the initial training features to the... n The distance between the center points; It is the first n One center point ( n = [1,2, , n ]); For the first i One memory feature; Let be the Euclidean distance function. Let be the set of all current memory features. A KD-Tree is constructed. This data structure recursively partitions the data space, which greatly accelerates the nearest neighbor search process. For each candidate center point... The mean shift vector is no longer calculated using a brute-force approach, but instead utilizes a KD-Tree for fast range lookup. In the... k In the next iteration, the search is performed to find the point that falls within the current center point. Centered on, with radius The mathematical representation of all neighboring points within the hypersphere is given by formula (8).

[0096] (8)

[0097] in, Based on the n The first center point k An index is generated, which contains the samples with the most similar features. The next step is to calculate the new center point location, which is the mean of all points in the neighborhood, and can be expressed by formula (9).

[0098] ; (9)

[0099] in, This indicates the updated center point. Indicates the first in the neighborhood i The process continues until the center point converges (i.e., the offset). (The number of iterations must be less than a set threshold) or the maximum number of iterations must be reached. The final set of stable center points is the density pattern center of the memory feature.

[0100] When new input features arrive, the new input features are calculated along with the features remembered at the current stage. mem The distance to the center point is calculated as shown in formula (10).

[0101] (10)

[0102] if , then this input feature ( ) will be deleted without any model updates. );if , then this input feature ( The model is updated according to the following rules. This update process only adjusts the connection weights of the neural network and does not increase the number of neurons. When a new input feature enters the training process, the information transfer variable ( The formula (11) needs to be calculated.

[0103] (11)

[0104] in, Based on the information transmission variables, the output weights at this stage can be updated using formula (12).

[0105] (12)

[0106] in, For the firsti The actual target value corresponding to the step. If Then, the number of neurons in the neural network is increased using formulas (3) - (6). Following the above training process, the updated output weights ( ) can be used to predict the corresponding target, which can be expressed by formula (13).

[0107] (13)

[0108] Finally, by iterating through all the training data, the updated set of weights will become the final output weights of the prediction neural network. Table 1 shows the training process of OSK-ESN-MFKD.

[0109] Table 1

[0110]

[0111] To verify the predictive ability of the proposed algorithm for water level, this embodiment uses artificial data and water level data to verify and compare the model and method.

[0112] The Lorenz data, an example of weather data representing chaotic systems, is a simplified version of the Navier-Stokes equations used in fluid mechanics. The Lorenz equations are written as follows:

[0113] ;

[0114] ;

[0115] ;

[0116] in, = 10, =8 / 3, =28. Consider the x-coordinate of the Lorenz time series for prediction and generate a time series of length 2536. In this artificial dataset, the prediction range (P) is set to eighteen steps, and the prediction time window (D) is eighteen in size.

[0117] In the experimental data involved in this invention, daily water level records from four representative water level observation stations in the Nanjing section of the Yangtze River (including Jiujiang, Hankou, Nanjing, and Maoping) were selected, with data collection time uniformly set at 08:00 daily. The data spans from May 15, 1984 to November 29, 2016. To verify the adaptability and robustness of the proposed model under different data scales, the data scale for each station was differentiated: a small sample set was used for Jiujiang and Hankou stations, a medium-sized sample set was used for Maoping station, and a large sample set was used for Nanjing station. Simultaneously, during training, the prediction range (P) was set to seven, and the time window (D) was set to seven. Based on these settings, and through data transformation methods, a data matrix for model training can be generated. Specific data configuration details are shown in Table 2.

[0118] Table 2

[0119]

[0120] On the other hand, to demonstrate the superior predictive performance of this invention and prove its effectiveness on water level data of different magnitudes, various existing online prediction models were selected for comparison, including: Online Sequential Extreme Learning Machine (OS-ELM), Online Sequential Echo State Network (OS-ESN) proposed in this invention, and Self-attention Weighted Online Sequential Echo State Network (SW-OSESN). Furthermore, all data were divided into training and testing datasets in different models at a ratio of 70% and 30%.

[0121] Meanwhile, to ensure all comparative models achieve optimal predictive performance, parameter selection for each model is a crucial consideration during training. In training all comparative models, a grid search method with intervals of 10 and a range of [10, 1000] was used to find the optimal number of hidden neurons. For OS-ESN and SW-OSESN, the spectral radius also needed to be determined. The optimal spectral radius was determined by searching with intervals of 0.05 and a range of [0.05, 0.95]. Furthermore, in the randomized algorithm, to ensure consistent predictive performance, the initialization seed was set to 1000 in the experiments.

[0122] Finally, to evaluate the predictive performance of the proposed method and prediction model across different orders of magnitude and time periods, Mean Square Error (MSE) and Symmetric Mean Absolute Percentage Error (SMAPE) were used as metrics to measure the model's predictive ability. The formulas are as follows:

[0123] ;

[0124] ;

[0125] in, Indicates the first i A true value, Indicates the first i One predicted value, n This indicates the total number of data points.

[0126] Table 3 presents a comparison of the performance of the online models (OS-ELM, OS-ESN, SW-OSESN) that currently perform well in water level prediction with the online prediction model OSK-ESN-MFKD proposed in this invention in terms of prediction performance on artificial data.

[0127] Table 3

[0128]

[0129] In the artificial dataset LO, the prediction results of the four models shown in Table 3 demonstrate that OSK-ESN-MFKD achieved the lowest MSE (7.88E-04) and SMAPE (5.99%) across the overall prediction period (1–18), indicating that this model has the best overall performance in long-term prediction. Compared with other models, OSK-ESN-MFKD also consistently demonstrated superior or highly competitive prediction accuracy in the short-term (1–6), medium-term (7–12), and long-term (13–18) sub-periods.

[0130] Specifically, in short-term forecasting, SW-OSESN had the lowest MSE, but OSK-ESN-MFKD's SMAPE was only 1.83%, outperforming all other models. In medium-term forecasting, OSK-ESN-MFKD's SMAPE was 6.35%, significantly lower than other models. In long-term forecasting, both its MSE and SMAPE were superior to other models except SW-OSESN. Overall, OSK-ESN-MFKD, compared to the second-best performing SW-OSESN, showed a significant performance improvement in long-term forecasting, with an MSE reduction of approximately 33.22% and a SMAPE reduction of 18.28%. Furthermore, results from different forecasting step sizes showed that short-term forecasting accuracy was generally higher than medium- and long-term forecasting accuracy, and the forecasting error gradually increased with the number of steps. To further illustrate the predictive capabilities of OSK-ESN-MFKD, Figure 2 The comparison between the predicted values ​​(blue curve) and actual values ​​(red curve) of the model in step seven of the Lorenz data is presented. It can be seen that the predicted curve and the actual curve are highly consistent in overall trend, demonstrating a good fit and confirming the model's strong accuracy and stability in complex time series forecasting.

[0131] In the comparative experiments on water level data, Table 4 shows the performance of the comparative model and the model proposed in this invention at different step counts and average periods for water level data of different scales. Experimental results on the datasets of four water level stations (JJ, HK, NJ, and MP) show that the OSK-ESN-MFKD model outperforms or matches the best existing model in most prediction step lengths and overall periods 1–7, especially demonstrating a more significant error control advantage at HK and MP stations. Specifically, at HK station, the model achieved the lowest MSE (6.87E-05) and SMAPE (4.37%) throughout the entire prediction process, with the smallest short-term and medium-term prediction errors; at MP station, all prediction step lengths and overall error indices were optimal; at JJ station, although the SMAPE was slightly higher than SW-OSESN at the seventh step, the overall SMAPE was still the lowest; at NJ station, the overall MSE was superior to the comparative model, and the SMAPE also performed well.

[0132] Table 4

[0133]

[0134] Furthermore, the comparison results of the seventh-step predicted values ​​and actual values ​​of OSK-ESN-MFKD at four water level stations ( Figure 3As can be seen, the predicted curve and the actual value are highly consistent with the trend of the entire time series, which further confirms that the model has good fitting accuracy and prediction stability for complex water level data.

[0135] In the above embodiments, a dynamically growing echo state network structure based on kernel methods is used for channel water level prediction. This structure uses kernel methods to construct reserve states, replacing traditional random weight initialization, thus improving feature representation ability and prediction stability. An online learning mechanism is employed to dynamically add hidden neurons based on input data, enabling the model to adaptively learn new patterns in water level changes. Simultaneously, this invention also proposes a network update and redundancy control mechanism based on unsupervised learning filters. This mechanism uses Mean-shift and KD-tree algorithms to calculate the distance between new data and existing knowledge centers in real time, intelligently deciding on neuron additions, weight updates, or data deletion; thereby achieving autonomous online model optimization while effectively preventing the network structure from expanding indefinitely.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting channel water levels based on an online kernel echo state network using unsupervised learning filters, characterized in that, The method includes: S1. Obtain water level data of the waterway recorded by multiple water level observation stations deployed along the waterway; S2. Construct and train an online kernel echo state network based on an unsupervised learning filter as a medium- to long-term channel water level prediction model; during model training and updates, the center point of the training reserve memory is found through Mean-shift clustering and KD-tree spatial indexing techniques. The relationship between the new data and the reserve memory features is determined by comparing the center point with the new data. Based on the determination result, it is determined whether to add new neurons to learn new features, update only the connection weights to fine-tune the model, or delete redundant data; the kernel method is used to calculate the reserve pool state in the echo state network; wherein, finding the center point of the training reserve memory through Mean-shift clustering and KD-tree spatial indexing techniques includes: The distance between a new input feature and existing features is obtained by calculating the minimum Euclidean distance between the new input feature and each cluster center, expressed as: ； in, Indicates the first i From the initial training features to the... n The distance between the center points; It is the first n One central point; For the first i One memory feature; It is the Euclidean distance function; For the current set of all memory features Construct a KD-Tree and recursively partition the data space; For each candidate center point Use KD-Tree for fast range queries; In the k In the next iteration, the search is performed to find the point that falls within the current center point. Centered on, with radius All neighboring points within the hypersphere are represented as: ； in, Based on the n The first center point k An index containing samples with the most similar features; The new center point location, i.e., the mean of all points in the neighborhood, is calculated as follows: ； in, Indicates the updated center point; Indicates the first in the neighborhood i The process continues until the center point position converges or the maximum number of iterations is reached; the final set of stable center points is the density pattern center of the memory feature. S3. Based on the acquired water level data, use the trained water level prediction model to predict the water level of the waterway.

2. The method for predicting channel water levels based on an online kernel echo state network using an unsupervised learning filter as described in claim 1, characterized in that, By comparing the center point with the newly added data, the relationship between the newly added data and the reserved memory features is determined. Based on the determination result, the model weights are updated, including: When there are new input features Upon entry, calculate the new input features and the features based on the current memory. mem Calculated center point Distance: ； if If so, then delete the new input feature. No model updates are performed; if Then, for this new input feature The model update process only adjusts the connection weights of the neural network without increasing the number of neurons; among them, It is the set distance threshold; if This increases the number of neurons in the neural network.

3. The method for predicting channel water levels based on an online kernel echo state network using an unsupervised learning filter, as described in claim 2, is characterized in that... Increase the number of neurons in the neural network and adjust the connection weights, including: The first nuclear stockpile status Calculate using the following formula: ； in, This represents the first input feature in the training data; ( ) is the Gaussian kernel function; This indicates the memory feature selected by the model; in this case, the memory feature is the first input feature. Based on intermediate variables Initialized output weights Use the following formula to perform the calculation: ； in, The target value corresponding to the first trained feature. To initialize intermediate variables: ； Based on the newly entered input features Nuclear stockpile status Calculation: ； in, It is the spectral radius coefficient. These are the memory features selected by the model, expressed as ; For intermediate variables Update: ； in, ; Calculate the error using the output weights from the previous step: ； The current output weight is calculated based on the error. Update: ； Through the above update process, whenever a new input feature enters the model, the predictive neural network will automatically add corresponding neurons according to the input feature, enabling the model to learn new predictive characteristics and patterns.

4. The method for predicting channel water levels based on an online kernel echo state network using an unsupervised learning filter, as described in claim 3, is characterized in that... When a new input feature is introduced into the training process, the information transfer variable is calculated. The calculation formula is: ； in, ; Based on the information transmission variables, the output weights at this stage are updated according to the following formula: ； in, For the first i The actual value of the target corresponding to each step.

5. The method for predicting channel water levels based on an online kernel echo state network using an unsupervised learning filter, as described in claim 4, is characterized in that... Predicting channel water levels using a trained channel water level prediction model includes: predicting channel water levels using updated output weights, expressed as: ； in, For the first The predicted water level of the waterway corresponding to each step; These are the updated output weights obtained after model training. It is an intermediate variable.

Images