An Artificial Intelligence Tide Prediction System and Method Based on Indoor Water Level and Air Pressure Data

Abstract

This invention discloses an artificial intelligence-based tide prediction system and method based on water level and air pressure data within an air chamber, belonging to the field of marine energy power generation technology, and applicable to oscillating water column wave energy power generation devices. The system includes: a data acquisition unit for simultaneously acquiring time-series water level and air pressure data within the air chamber of the device; a data preprocessing unit for preprocessing the water level and air pressure time-series data to obtain effective features; and an artificial intelligence tide prediction model unit, including an artificial intelligence tide prediction model employing a time-series convolutional network customized and improved for the oscillating water column device scenario. This model extracts the fluid-structure interaction spatial features of water level and air pressure, as well as the time features of low-frequency tidal trends, based on the effective features, and performs spatiotemporal fusion, while simultaneously capturing long-period temporal dependencies of tides and outputting predicted tide values. This invention enables stable, accurate, and real-time prediction of tide levels under various sea conditions.

Description

Technical Field

[0001] This invention relates to the field of ocean energy power generation technology, specifically to an artificial intelligence tide level prediction system and method based on water level and air pressure data in an air chamber. Background Technology

[0002] With the global energy structure shifting towards cleaner energy sources, wave energy, as a abundant and renewable clean marine energy source, has become a research hotspot in the field of ocean power generation. Oscillating water column (OWC) wave energy generation devices have become one of the most widely used wave energy generation technologies due to their advantages such as simple structure, high reliability, and wide adaptability to various sea conditions.

[0003] During the operation of an OWC (Overflow-Cooled Gas) unit, tidal level is a key parameter affecting the unit's hydrodynamic performance, structural safety, and power generation efficiency. Changes in tidal level directly determine the connection depth between the gas chamber and the external sea area, thus influencing the oscillation amplitude of the water column within the gas chamber under wave action, the pattern of air pressure changes, and the turbine's working efficiency. For example, during spring tides, the effective water depth of the gas chamber increases, improving the energy capture efficiency of the water column oscillation, but also increasing the wave impact force on the unit structure. During neap tides, the effective water depth decreases, reducing energy capture efficiency, requiring compensation through adjustments to turbine operating parameters. Therefore, real-time and accurate tidal level information is the core data support for OWC unit control optimization, efficiency assessment, and maintenance decisions.

[0004] However, existing tide level measurement and prediction technologies have significant shortcomings in the application scenarios of OWC devices, as follows: 1. Traditional measurement methods suffer from low accuracy and poor anti-interference capabilities: Traditional tide level observations largely rely on shore-based tide gauges or radar tide meters. Shore-based tide gauges are limited by geographical conditions and can only be deployed in near-shore areas, failing to provide effective tide level data for offshore OWC devices far from the shoreline. Furthermore, setting up dedicated tide gauges is costly and unsuitable for temporary projects or distributed wave energy farms. While radar tide meters can be directly installed on the device, large waves and violent surface fluctuations severely interfere with measurements, leading to data anomalies, distortions, or even failures. Simultaneously, radar tide meters are susceptible to environmental factors such as wave splash, fog, and rain, exhibiting extremely poor stability in harsh sea conditions. In addition, contact-type measurement devices such as pressure tide meters are prone to marine organism adhesion and seawater corrosion, resulting in high maintenance costs, and they also cannot avoid measurement noise caused by short-period wave fluctuations.

[0005] 2. Existing AI prediction models are not adapted to the characteristics of OWC scenarios: In existing technologies, some research attempts to predict tide levels using AI algorithms, but most of them use general-purpose Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), or primitive Temporal Convolutional Networks (TCNs), which are not adapted to the specific scenarios of OWC devices, resulting in the following problems: ① Inability to handle the fluid-structure interaction between the inside and outside of the air chamber: The changes in water level and air pressure inside the OWC device's air chamber are the result of the combined effects of waves, tides, and the air chamber structure, involving a complex coupling process of fluid dynamics and aerodynamics—wave impacts cause instantaneous oscillations in water level, which in turn trigger pulse-like changes in air pressure, while tides modulate the baseline of this oscillation by altering the effective water depth. Traditional AI models can only process single-dimensional time-series data and cannot effectively explore the spatiotemporal coupling relationship of "instantaneous response-baseline modulation" between water level and air pressure; ② Insufficient multi-scale signal separation capability: The water level fluctuation in the air chamber contains two components—short-period (second-level) wave oscillation noise and long-period (hour-level) tidal trend signal. The energy difference between the two can reach 1 to 2 orders of magnitude. Traditional models have difficulty effectively separating these two scale signals, resulting in wave noise submerging tidal features under severe sea conditions and a sharp drop in prediction accuracy. ③ Low efficiency in capturing long-term time-series dependencies: Tidal changes have significant long-period characteristics (e.g., a semi-diurnal tide cycle of about 12.42 hours and a diurnal tide cycle of about 24.84 hours). Traditional LSTM / GRU models suffer from gradient vanishing problems when processing long-term time-series data, and have weak parallel computing capabilities, resulting in low real-time inference efficiency. Although the original TCN has parallel advantages, its inflation factor increases exponentially with the number of network layers, leading to excessive zero-filling in deep networks and dilution of long-period tidal features, making it impossible to accurately capture the periodic dependencies of tides.

[0006] 3. Limitations of Existing Technologies: In recent years, some scholars have proposed improved TCN architectures, but such technologies are currently only applied to fields such as remaining life prediction for aero-engines. Their core objective is to capture the nonlinear degradation process of equipment "from health to failure," with input being performance degradation data from multiple sensors and output being remaining life duration (single-dimensional decreasing labels). In contrast, the core requirement for OWC device tidal level prediction is "extracting stable low-frequency trends from high-frequency oscillation signals," with input being fluid-structure interaction data of water level and air pressure, and output being continuously changing tidal level elevation (periodic continuous labels). The application scenarios, data characteristics, and core pain points of the two are completely different. Existing technologies have not been adapted to the specific requirements of OWC scenarios such as "fluid-structure interaction, multi-scale signal separation, and tidal periodic capture," making direct application impossible. Furthermore, they have not addressed industry pain points such as wave noise interference and long-period tidal signal extraction in tidal level prediction.

[0007] In summary, existing technologies cannot meet the application requirements of OWC wave energy generation devices for "high precision, strong anti-interference, real-time online, and low cost" of tide level data. There is an urgent need for an artificial intelligence tide level prediction technology based on the easily monitorable parameters of the OWC device and its deep adaptation to various scenarios. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide an artificial intelligence-based tide prediction system and method based on water level and air pressure data within an air chamber. This system solves the problems of low accuracy, poor anti-interference capability, and reliance on external specialized equipment in traditional tide measurement methods under wave environments. It also overcomes the limitations of existing AI models, such as inability to adapt to the fluid-structure interaction characteristics of open-water environments (OWC), insufficient multi-scale signal separation capability, and low long-term-series-dependent capture efficiency. The present invention utilizes easily monitorable internal parameters of the OWC device (liquid level height and air pressure changes within the air chamber) and establishes a nonlinear mapping model from internal parameters to external tide levels through a customized and improved artificial intelligence algorithm tailored to the physical characteristics of OWC scenarios. This enables stable, accurate, and real-time prediction of tide levels under various sea conditions, providing reliable data support for OWC device control optimization, efficiency evaluation, and maintenance decisions.

[0009] This invention provides an artificial intelligence tide prediction system based on water level and air pressure data in an air chamber, applied to an oscillating water column wave energy power generation device. The system includes: The data acquisition unit is used to simultaneously acquire the water level time-series data and air pressure time-series data in the air chamber of the device; A data preprocessing unit, connected to the data acquisition unit, is used to preprocess the water level time series data and the air pressure time series data. The preprocessing includes at least: constructing multiple candidate features and selecting effective features from the multiple candidate features based on the mapping contribution between each candidate feature and the external tide level label. The artificial intelligence tide prediction model unit, connected to the data preprocessing unit, includes an artificial intelligence tide prediction model that employs a temporal convolutional network customized and improved for the oscillating water column device scenario. The temporal convolutional network is used to extract the fluid-structure interaction spatial features of water level and air pressure and the temporal features of low-frequency tidal trends based on the effective features and complete the spatiotemporal fusion. At the same time, it captures the long-period temporal dependence of tides and outputs the tide prediction value.

[0010] Preferably, the data acquisition unit includes a chamber water level sensing module, a chamber air pressure sensing module, and a synchronous acquisition control module; The air chamber water level sensing module is an immersion-type liquid level gauge used to collect time-series data of the water level inside the air chamber; The air chamber pressure sensing module is a high-precision pressure transmitter with the same sampling frequency as the air chamber water level sensing module, used to collect time-series air pressure data inside the air chamber. The synchronous acquisition and control module performs synchronous management and control of the submersible level gauge and the high-precision pressure transmitter.

[0011] Preferably, the data preprocessing unit includes a loss boundary-mapping capability feature filtering module, which is used for: Determine the loss boundary value, which is the minimum theoretical loss value that a preset single-feature benchmark regression model can achieve when inputting random noise; For each candidate feature, determine the actual loss value for that candidate feature. The actual loss value is the loss value obtained after training the preset single-feature benchmark regression model when the candidate feature is used as the only input. Based on the loss boundary value and the actual loss value, calculate the mapping capability quantization value of the candidate feature; Candidate features with a mapping capability quantization value greater than zero are retained as valid features.

[0012] Preferably, the formula for calculating the loss boundary value in the loss boundary-mapping capability feature filtering module is as follows: ; In the formula, This represents the boundary value of the loss. This represents the total number of samples in the pre-training dataset; For the first Random noise prediction value for each sample; For the first The actual external tide level label value for each sample; For the first in the dataset One sample.

[0013] Preferably, the data preprocessing unit includes a data cleaning module, a data normalization module, and a time-series feature construction module; The data cleaning module uses a three-standard-deviation criterion to remove outliers, linear interpolation to fill missing data, and a five-step moving average filtering to remove noise. The time-series feature construction module constructs a multi-dimensional feature set including historical water level / air pressure values, water level change rate and air pressure change rate per unit time, cross-product features, and statistical features.

[0014] Preferably, the customized improved temporal convolutional network includes: The feature extraction layer is used to perform spatiotemporal coupling processing on the effective features. It extracts the fluid-structure interaction spatial features of water level and air pressure through spatial dimension channel selection, filters high-frequency wave noise and retains low-frequency tidal trend time features through time dimension gated convolution, and outputs the fused features after fusing the fluid-structure interaction spatial features and low-frequency tidal trend time features. Multiple temporal convolutional hidden layers are connected to the feature extraction layer to perform temporal convolution processing on the fused features in order to capture the long-period temporal dependencies of tides and output deep temporal features. A fully connected output layer, connected to the temporal convolutional hidden layer, is used to map the deep temporal features into a single-dimensional external tide level prediction value.

[0015] Preferably, the feature extraction layer includes: The spatial dimension channel selection unit is used to dynamically adjust the weight of each feature channel according to the fluid-structure interaction characteristics of water level and air pressure, strengthen the feature channels related to tides and suppress the wave noise channel, so as to output the fluid-structure interaction spatial characteristics. The time-dimensional gated convolutional unit is used to filter high-frequency wave noise through the reset gate and retain low-frequency tidal trend features through the update gate, so as to output low-frequency tidal trend time features. The dual-path feature fusion unit is used to fuse the fluid-structure interaction spatial features output by the spatial dimension channel selection unit with the low-frequency tidal trend temporal features output by the temporal dimension gated convolution unit to obtain fused features. The fused features are then processed by batch normalization and activation functions and used as the output of the feature extraction layer.

[0016] Preferably, the plurality of temporal convolutional hidden layers are configured using a segmented re-dilation factor strategy; in the segmented re-dilation factor strategy, in at least one pair of adjacent temporal convolutional hidden layers, the dilation factor of the later layer is less than or equal to the dilation factor of the earlier layer.

[0017] Preferably, the system further includes a model online adaptive update unit and a tide level output and application interface unit, wherein the model online adaptive update unit includes: The model performance monitoring module is used to monitor the prediction performance of the artificial intelligence tide level prediction model unit and determine whether the prediction performance is lower than a preset threshold. An online incremental training module is used to perform incremental fine-tuning training on the artificial intelligence tide level prediction model unit based on newly added on-site collected data when the prediction performance is lower than a preset threshold, so as to update the model parameters. The tide level output and application docking unit realizes the output and storage of tide level prediction values ​​and docking with the engineering application of the oscillating water column wave energy power generation device.

[0018] This invention also provides an artificial intelligence-based tide prediction method based on water level and air pressure data in an air chamber, applied to the aforementioned system, comprising the following steps: Step S1: Construct a full-scene dataset containing water level / pressure data of the air chamber of the oscillating water column wave energy power generation device and corresponding tide level labels, complete the offline pre-training of the artificial intelligence tide level prediction model, solidify the model parameters, and obtain the trained artificial intelligence tide level prediction model. Step S2: Synchronously collect water level and air pressure time series data in the air chamber of the oscillating water column wave energy generator. The water level and air pressure time series data inside the air chamber of the oscillating water column wave energy generator are collected synchronously through the data acquisition unit. Step S3: Input the water level time series data and air pressure time series data into the data preprocessing unit, and perform cleaning, normalization and time series feature construction in sequence. Then, calculate the feature mapping capability through the loss boundary-mapping capability feature screening module and retain the effective features to obtain the effective feature set. Step S4: Input the effective feature set into the trained artificial intelligence tide prediction model for inference and prediction, and output the tide prediction value; Step S5: The predicted tide level is denormalized to restore the actual physical quantity of the tide level. The actual physical quantity of the tide level is compared with the data from the tide gauge station every day. When the average absolute error exceeds ±5cm for 3 consecutive days, the artificial intelligence tide level prediction model is adjusted by constructing an incremental training set with the newly added data and the original artificial intelligence tide level prediction model is replaced, so as to realize the online adaptive update of the model.

[0019] The artificial intelligence tide prediction system and method based on water level and air pressure data in an air chamber, as disclosed in this invention, have the following advantages compared to existing technologies: 1. Targeted solutions to existing pain points: The loss boundary-mapping capability feature screening module is customized for the tide prediction scenario. By quantifying the mapping capability between features and tide labels, it separates wave noise and tidal features from the input source, solving the problem that existing technologies cannot effectively suppress interference from severe sea conditions; The customized temporal convolutional network architecture is fundamentally modified for the physical characteristics of the OWC scenario, rather than a simple domain migration: It adapts to the dynamic weight changes of water level-air pressure fluid-structure interaction, and the segmented resetting expansion factor strategy adapts to the signal separation requirements of long tidal periods and short wave periods, solving the inapplicability of existing algorithms in the OWC scenario; 2. High precision and strong anti-interference capability: Even under extreme high wave sea conditions (H s The predicted MAE (>4m) is ≤ ±4.2cm, which is 74.5% lower than the traditional LSTM model and 61.1% lower than the original temporal convolutional network model, solving the problem of the failure of traditional measurement methods under severe sea conditions; 3. Reduced system cost and complexity: No additional external dedicated tide level measurement equipment is required, reducing the cost by 60% compared to traditional radar tide level gauge solutions, and reducing installation complexity and potential failure points; 4. Strong real-time online prediction capability: single sample inference time ≤ 5ms, supports local edge deployment, and meets the real-time control requirements of OWC devices; 5. Excellent long-term stability: Through the online incremental update mechanism, the MAE remains ≤±3.4cm after 6 months of continuous operation, adapting to changes in marine characteristics and device performance. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of an artificial intelligence tide prediction system based on water level and air pressure data in an air chamber, according to the present invention.

[0021] Figure 2 This is a schematic diagram of the oscillating water column type OWC wave energy generation device of the present invention.

[0022] Figure 3 This is a flowchart of an artificial intelligence tide level prediction method based on water level and air pressure data in an air chamber, according to the present invention. Detailed Implementation

[0023] The following detailed description of an artificial intelligence tide prediction system and method based on water level and air pressure data in an air chamber, in conjunction with specific embodiments, illustrates the present invention. These embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0024] Example 1: Implementation of an artificial intelligence tide prediction system based on water level and air pressure data in an air chamber.

[0025] Combined with appendix Figures 1-2 As shown, the present invention provides an artificial intelligence tide prediction system based on water level and air pressure data in an air chamber.

[0026] like Figure 2 As shown, this system is mounted on an oscillating water column type OWC wave energy power generation device. The OWC wave energy power generation device includes several wave energy power generation structural units arranged in an array. Each wave energy power generation structural unit 1 includes a lower opening 11 that communicates with the external seawater, an air chamber body 12 with a closed air pressure chamber at the top, and a turbine power generation unit 13 that communicates with the air pressure chamber. External waves act on the water surface inside the air chamber, causing the water column inside the air chamber to oscillate up and down and drive changes in air pressure, thereby driving the turbine power generation unit to generate electricity. Figure 2 The markings in the text are for illustrative purposes only, indicating the device structure and connection relationships. This system specifically includes a data acquisition unit 110, a data preprocessing unit 120, an artificial intelligence tide level prediction model unit 130, a model online adaptive update unit 140, and a tide level output and application interface unit 150. Each unit is electrically / communicationally connected in sequence.

[0027] 1. Data acquisition unit 110 The data acquisition unit 110 is used to realize the synchronous and high-fidelity acquisition of water level and air pressure data inside the air chamber of the OWC device. The acquired water level time series data and air pressure time series data inside the air chamber of the device are the data source foundation of the system.

[0028] The data acquisition unit 110 includes a chamber water level sensing module 111, a chamber air pressure sensing module 112, and a synchronous acquisition control module 113. The chamber water level sensing module 111 is a submersible level gauge used to acquire time-series data of the water level inside the chamber. The chamber air pressure sensing module 112 is a high-precision air pressure transmitter with the same sampling frequency as the chamber water level sensing module, used to acquire time-series data of the air pressure inside the chamber. The synchronous acquisition control module 113 performs synchronized control of the submersible level gauge and the high-precision air pressure transmitter. This unit specifically includes: (1) Air chamber water level sensing module 111: It adopts an immersion liquid level gauge and is installed in the area without direct wave impact on the inner side wall of the air chamber of the OWC device (at 1 / 3 of the height from the bottom of the air chamber). The measurement range is 0~5m, the measurement accuracy is ±1cm, and the sampling frequency can be configured to 5~20Hz to meet the sampling requirements of long period tidal signals and short period wave signals. (2) Air chamber pressure sensing module 112: A high-precision pressure transmitter is used and installed in the closed pressure chamber at the top of the air chamber (away from the turbine inlet to avoid the influence of airflow disturbance). The measurement range is 80~120kPa, the measurement accuracy is 0.1%FS, and the sampling frequency is completely consistent with the water level sensing module. (3) Synchronous acquisition control module 113: It adopts an STM32H7 series microcontroller and is electrically connected to the air chamber water level sensing module 111 and the air chamber air pressure sensing module 112 via a 485 bus. Its core function is to control the synchronous sampling of the two sensors (timestamp error ≤10ms) to ensure the time sequence consistency of water level and air pressure data; at the same time, it completes the temporary storage and preliminary verification of the raw data, removes abnormal values ​​that are obviously beyond the measurement range, and transmits the raw time sequence data to the data preprocessing unit 120 via Ethernet or LoRa wireless communication.

[0029] 2. Data preprocessing unit 120 The data preprocessing unit 120 is used to clean, standardize and feature-engineer the raw collected data, eliminate data noise and the influence of dimensions, and construct a feature set that is adapted to the input requirements of artificial intelligence models.

[0030] The data preprocessing unit 120, connected to the data acquisition unit 110, is used to preprocess the water level time series data and the air pressure time series data. The preprocessing includes at least: constructing multiple candidate features and selecting effective features from the multiple candidate features based on the mapping contribution between each candidate feature and the external tide level label.

[0031] The data preprocessing unit 120 includes a loss boundary-mapping capability feature filtering module 124, which is used for: Determine a loss boundary value, which is the minimum theoretical loss value that a preset single-feature baseline regression model can achieve when inputting random noise; for each candidate feature, determine the actual loss value of that candidate feature, which is the loss value obtained after training the preset single-feature baseline regression model when using only that candidate feature as the sole input; calculate the mapping capability quantization value of the candidate feature based on the loss boundary value and the actual loss value; retain candidate features with a mapping capability quantization value greater than zero as the effective features.

[0032] The data preprocessing unit 120 also includes a data cleaning module 121, a data normalization module 122, and a time-series feature construction module 123. The data cleaning module 121 uses a three-standard-deviation criterion to remove outliers, linear interpolation to fill missing data, and a 5-step moving average filtering to remove noise. The time-series feature construction module 123 constructs a multi-dimensional feature set containing historical water level / air pressure values, water level change rate and air pressure change rate per unit time, cross-product features, and statistical features. This unit specifically includes: (1) Data cleaning module 121: preprocesses and reduces noise in the original water level and air pressure time series data, using 3D printing technology. The criteria (i.e., the three-standard-deviation criterion) are used to eliminate abnormal data caused by instantaneous sensor jitter and wave impact; linear interpolation is used to complete short-term (≤5 sampling steps) missing data; and a 5-step moving average filter is used to remove high-frequency acquisition noise and retain effective fluctuation signals related to tides and waves. (2) Data normalization module 122: The cleaned water level and air pressure data are normalized to the [0, 1] interval using Min-Max normalization to eliminate the influence of different parameter dimensions on model training and inference. The normalization formula is: ; In the formula: To normalize the data, The original data, These are the minimum and maximum values ​​of the dataset, respectively. A small constant is added to the denominator to avoid division by zero errors. (3) Temporal feature construction module 123: Based on the normalized temporal data, construct a multi-dimensional input feature set, including: the water level and air pressure values ​​of the air chamber at the current time and the historical M times (M=500 sampling steps); the water level change rate and air pressure change rate per unit time; the water level-air pressure cross-product feature (characterizing the fluid-structure interaction strength); the water level statistical features (mean, variance, peak value, valley value) and air pressure statistical features within a 30-step sliding window; (4) Loss Boundary to Mapping Ability Feature Filtering Module 124: Addressing the pain point of strong coupling between wave noise and tidal features in OWC scenarios, this module uses a customized Loss Boundary to Mapping Ability (LM) method to quantify the mapping contribution of each feature to the tide level label, filter effective features, and eliminate noisy features. This module is the core support for the inventiveness of this invention. Its essential difference from existing technologies lies in redefining the calculation logic of the loss boundary and mapping ability for the periodic and continuous label characteristics of tide level prediction, rather than simply reusing general feature filtering methods. The specific implementation is as follows: ① Based on the "minimum theoretical loss when there are no effective features", the ability of a single feature to be mapped to the tide label through the benchmark model is quantified, and only features with positive mapping ability are retained, thus separating wave noise and tidal features from the input source.

[0033] ②Loss boundary value When the input features have no effective mapping relationship with the tide level label, the minimum theoretical loss value that the baseline model can achieve is used as the benchmark for evaluating the feature mapping capability. The calculation formula is as follows: ; In the formula, This represents the boundary value of the loss. This represents the total number of samples in the pre-training dataset; For the first Random noise prediction value for each sample; For the first The actual external tide level label value for each sample; For the first in the dataset One sample; This refers to the minimum theoretical loss value that the benchmark model can achieve when there is no effective mapping relationship between the input features and the tide level label, and serves as a benchmark for evaluating the feature mapping capability. The total number of OWC device air chamber sensing data and corresponding tide level label samples covers different sea conditions, different tidal cycles, and different seasons. When the input to the baseline model is random noise unrelated to the tide level label, the output is the first... Predicted tide levels for each sample; Data was collected using a high-precision pressure tide gauge synchronized with GPS time of the air chamber sensor. The summation function calculates the summation of elements from the first to the last element in the pre-training dataset. The corresponding items of each sample are summed up for calculation; : No. The squared difference between the predicted random noise value and the actual tide level label value of each sample is used to quantify the prediction bias. Mean squared error (MSE) is calculated as the average of the squared differences of all samples. The minimum value function takes the minimum RMSE (Recovery Mean Squared Error) of all possible outputs when the baseline model is input with random noise; this is the loss boundary value. .

[0034] ③ Actual loss value of a single feature Regarding the first For each feature to be evaluated, this single feature is used as the sole input and trained under supervision using a pre-defined baseline regression model (a 2-layer fully connected network with 1 input dimension, 16 hidden layer dimensions, ReLU activation function, and 1 output dimension). The average validation loss over the last 10 epochs is taken as the actual loss for that feature, calculated using the following formula: ; In the formula: For only the first input When the first feature is reached, the output of the benchmark regression model is the first feature. Predicted tide levels for each sample; For the first The actual tide level label corresponding to each sample; The number of epochs used to train the baseline regression model; For the first The actual loss value of the first feature to be evaluated refers to the loss value of only the first feature. When a feature is used as input, the average validation loss value obtained after training the baseline regression model is used to quantify the basic contribution of the feature to tide level prediction. The serial number of the feature to be evaluated is a unique identifier corresponding to the air chamber water level, air pressure, and their derived features (such as water level change rate, air pressure statistical features, etc.). This represents the total number of samples in the pre-training dataset; For the first The first feature corresponds to the first The squared difference between the predicted value and the actual tide level label value for each sample; For the first The first feature corresponds to the first The mean square error (MSE) between the predicted value and the actual tide level label value for each sample.

[0035] ④ Feature mapping capability : Calculate the first loss by comparing the loss boundary value with the actual loss value of a single feature. The mapping ability of each feature to the tide level label is calculated using the following formula: ; In the formula: For the first The quantized value of the mapping capability of the feature is used to characterize the feature. The contribution of each feature to tide level prediction The larger the value, the greater the contribution of this feature to tide level prediction and the stronger its mapping ability; ReLU To ensure the mapping capability is non-negative for the linear rectifier function, when When this occurs, it means that the feature does not contribute effectively to tide level prediction and belongs to the invalid noise feature caused by wave interference. The loss boundary value serves as a benchmark for evaluating feature mapping capability. For the first The actual loss value of each feature to be evaluated; : Loss difference, when When, explain the first These features contribute effectively to tide level prediction; when If this occurs, it indicates that the feature makes no effective contribution; Normalized loss difference, mapping the loss difference to An interval is used to uniformly quantify the mapping strength of different features.

[0036] ⑤ Execution process: Calculate all features based on the pre-trained dataset. ,reserve Effective features; the model is re-screened quarterly during online operation to dynamically adapt to changes in sea conditions and device performance.

[0037] 3. Artificial Intelligence Tide Level Prediction Model Unit 130 The AI ​​tide prediction model unit 130 is the core functional unit of the system. It adopts a TCN backbone network (instantaneous convolutional network backbone network) that is customized and improved for the physical characteristics of the OWC scene, rather than simply reusing the existing technical architecture. The core improvements are centered around the unique requirements of the OWC device for fluid-structure interaction, multi-scale signal separation, and long-period tidal capture.

[0038] The artificial intelligence tide prediction model unit 130 is connected to the data preprocessing unit 120 and includes an artificial intelligence tide prediction model. It adopts a temporal convolutional network customized and improved for the oscillating water column device scenario. The temporal convolutional network is used to extract the fluid-structure interaction spatial features of water level and air pressure and the temporal features of low-frequency tidal trend based on the effective features and complete the spatiotemporal fusion. At the same time, it captures the long-period temporal dependence of tides and outputs the tide prediction value.

[0039] Customized and improved temporal convolutional networks include: The feature extraction layer is used to perform spatiotemporal coupling processing on the effective features. It extracts the fluid-structure interaction spatial features of water level and air pressure through spatial dimension channel selection, filters high-frequency wave noise and retains low-frequency tidal trend time features through time dimension gated convolution, and outputs the fused features after fusing the fluid-structure interaction spatial features and low-frequency tidal trend time features. Multiple temporal convolutional hidden layers are connected to the feature extraction layer to perform temporal convolution processing on the fused features in order to capture the long-period temporal dependencies of tides and output deep temporal features. A fully connected output layer, connected to the temporal convolutional hidden layer, is used to map the deep temporal features into a single-dimensional external tide level prediction value.

[0040] The feature extraction layer includes: The spatial dimension channel selection unit is used to dynamically adjust the weight of each feature channel according to the fluid-structure interaction characteristics of water level and air pressure, strengthen the feature channels related to tides and suppress the wave noise channel, so as to output the fluid-structure interaction spatial characteristics. The time-dimensional gated convolutional unit is used to filter high-frequency wave noise through the reset gate and retain low-frequency tidal trend features through the update gate, so as to output low-frequency tidal trend time features. The dual-path feature fusion unit is used to fuse the fluid-structure interaction spatial features output by the spatial dimension channel selection unit with the low-frequency tidal trend temporal features output by the temporal dimension gated convolution unit to obtain fused features. The fused features are then processed by batch normalization and activation functions and used as the output of the feature extraction layer.

[0041] In this unit, multiple temporal convolutional hidden layers are configured using a segmented re-dilation factor strategy. In this strategy, in at least one pair of adjacent temporal convolutional hidden layers, the dilation factor of the later layer is less than or equal to the dilation factor of the earlier layer. Specifically, this unit includes: (1) Model Architecture Module 131 ① Input layer: Receives the effective temporal feature set after LM feature filtering. The input feature tensor format is as follows: ( For batch size, For historical time series length, (Effective feature dimension).

[0042] ② Regarding the feature extraction layer for fluid-structure interaction: In existing technologies, the feature extraction layer is only used to extract general spatiotemporal features from multiple sensors. This invention fundamentally modifies the function and structure of the OWC device based on the physical characteristics of "water level-air pressure fluid-structure interaction". (a1) Spatial dimension channel selection unit: In view of the characteristic of "dynamic change of the contribution weight of water level and air pressure to tide level under different sea conditions", the feature channel strongly related to tide is strengthened by adaptive weight allocation (such as strengthening the average air pressure feature under large wave sea conditions and strengthening the peak water level feature under small wave sea conditions) and suppressing the wave noise channel. (a2) Time-dimensional gated convolutional unit: In view of the coexistence of short-period wave oscillations and long-period tidal trends, the reset gate filters out high-frequency wave noise at the second level, and the update gate retains the tidal trend characteristics at the hour level, thus solving the problem of multi-scale signal separation. (a3) Dual-path feature fusion: Spatial coupling features and temporal trend features are concatenated and then batch normalized to ensure the effectiveness of feature fusion. The calculation formula is as follows: ; In the formula: As a feature of fusion, Output spatial dimension features. This outputs time-dimensional features; Concat is the concatenation operation. BN For batch normalization operations, This is the activation function.

[0043] ③ TCN Hidden Layer (Customized Improvement for Long-Period Tidal Patterns): In existing technologies, the expansion factor only grows exponentially. This solution addresses the matching requirements between long-period tidal characteristics and OWC sampling frequency (5~20Hz) by designing a segmented expansion factor reset strategy. This solves the pain points of excessive zero-filling and dilution of tidal features in traditional deep TCN networks. The specific configuration is as follows: (b1) Two-layer configuration: The first layer has an expansion factor d=1 (the receptive field covers 10-second wave signals), and the second layer has an expansion factor d=2 (the receptive field covers 30-second signals), ensuring that short-cycle noise filtering and medium-cycle trend capture are taken into account. (b2) 3-layer configuration (preferred): Layer 1 d=1, Layer 2 d=2, Layer 3 d=1. The first two layers rapidly expand the receptive field to cover the long tidal period, and the third layer resets the expansion factor to reduce zero filling and avoid feature dilution. (b3) 4-layer configuration: Layer 1 d=1, Layer 2 d=2, Layer 3 d=4, Layer 4 d=2. The first three layers gradually expand the receptive field to the 2-hour level (covering key segments of the tidal cycle), and the fourth layer resets the expansion factor to balance long-term capture and feature concentration. (b4) Causal convolution constraint: Causal convolution is used throughout the process to ensure that the output at the current moment depends only on historical and current data, with no leakage of future information, which is suitable for real-time prediction needs; (b5) Residual block design: Each hidden layer contains improved dilated causal convolution + weight normalization + ReLU + Dropout, and residual edge dimension matching is achieved through 1×1 convolution to avoid gradient vanishing in deep networks; ④ Dropout layer (i.e., random deactivation layer): The random deactivation ratio is 0.1~0.3, which suppresses overfitting; ⑤ Fully connected output layer: The fused features are mapped to single-dimensional tide level predictions using a linear activation function.

[0044] (2) Offline pre-training module 132 ① Construction of pre-trained dataset: At a near-shore OWC test power station (latitude) ,longitude (semi-diurnal tides), collecting data on different sea conditions (small waves) Mid-wave Big waves Extreme sea conditions The data includes time-series data of air chamber water level and air pressure during different tidal cycles (spring tides and neap tides) and different seasons as input samples; a high-precision pressure tide gauge with GPS time synchronization is used simultaneously. Collect labeled samples and construct a full-scene dataset (100,000 records), which is then divided into training, validation, and test sets in a 7:2:1 ratio. ② Hyperparameter configuration: Feature extraction layer convolution kernel size 3, TCN hidden layer preferred configuration of 3 layers ( Dropout failure rate 0.2%, batch size 256, training 100 epochs, Adam optimizer (initial learning rate 0.001, cosine annealing decay). ③ Training and solidification: Using MSE as the loss function, an early stopping mechanism is set (the validation set loss decreases after 10 consecutive rounds). When the test set MAE (mean absolute error) is ≤ ±3cm, the model parameters are solidified.

[0045] (3) Real-time inference module 133 ① Model Deployment: Deployed on the local edge controller of the OWC device (NVIDIA Jetson Xavier NX), with a single-sample inference time of ≤5ms, meeting the real-time control requirements; ②Inference process: Input the real-time feature set filtered by LM, output the current tide level prediction value through forward inference, and restore it to the actual physical quantity of tide level through inverse normalization; supports outputting the tide level prediction sequence for the next 1 to 72 hours.

[0046] 4. Model online adaptive update unit 140 This unit is used to achieve long-term adaptive optimization of the artificial intelligence tide prediction model, and to solve the problem of decreased prediction accuracy caused by changes in sea tidal characteristics, seasonal sea state changes, and device performance degradation.

[0047] The online adaptive update unit 140 includes: a model performance monitoring module 141, used to monitor the prediction performance of the artificial intelligence tide prediction model unit 130 and determine whether the prediction performance is lower than a preset threshold; The online incremental training module 142 is used to incrementally fine-tune the artificial intelligence tide prediction model unit based on newly collected field data when the prediction performance is lower than a preset threshold, so as to update the model parameters. This unit specifically includes: ① Model performance monitoring module 141: The actual physical quantity of the tide level after daily inverse normalization is compared with the calibration benchmark data (synchronous data from high-precision tide gauge stations in the same sea area) to calculate the mean absolute error (MAE) and root mean square error (RMSE); when the MAE exceeds ±5cm for 3 consecutive days, the model update process is automatically triggered. ② Online Incremental Training Module 142: Collects newly added on-site air chamber water level and air pressure time series data and corresponding tide level annotation data, constructs an incremental training set (size is 20% of the original training set), performs incremental fine-tuning training on the basis of the original benchmark model parameters (30 training rounds, learning rate 0.0005), updates the model weight parameters, and completes the updated model to replace the original benchmark model for subsequent tide level prediction.

[0048] 5. Tide level output and application docking unit 150 The tide level output and application docking unit 150 is used to realize the output, storage and engineering application docking of tide level prediction values. It transmits the tide level prediction values ​​output by the model (which are restored to the actual physical quantity after inverse normalization) to the local controller and remote monitoring platform of the OWC wave energy power generation device in real time via Modbus TCP protocol. It also has the functions of local encrypted storage of tide level data (storage period of 90 days), historical backtracking query, and tide level anomaly alarm (triggered when the predicted tide level exceeds the preset safety range). It provides core tide level data support for the turbine start-up and shutdown control, valve opening adjustment, power generation prediction and operation and maintenance decision-making of the device.

[0049] Example 2: Implementation of an artificial intelligence-based tide prediction method based on water level and air pressure data in an air chamber.

[0050] Combined with appendix Figure 3 As shown, this invention provides an artificial intelligence-based tide level prediction method based on water level and air pressure data in an air chamber.

[0051] An artificial intelligence-based tide prediction method based on water level and air pressure data in an air chamber, applied to the aforementioned system, includes the following steps: Step S1: Construct a full-scene dataset containing water level / pressure data from the air chamber of an oscillating water column wave energy generator and corresponding tide level labels; complete the offline pre-training of the artificial intelligence tide level prediction model; solidify the model parameters; and obtain the trained artificial intelligence tide level prediction model. Step S1 specifically includes: S11: Construct a pre-trained labeled dataset: Collect time-series data of air chamber water level and air pressure under different sea conditions, different tidal cycles, and different seasons as input samples, and simultaneously collect high-precision tide gauge data as label samples to construct a full-scene dataset and divide it into training set, validation set, and test set. S12: Build a customized and improved TCN backbone network: Configure the parameters of each layer according to the above architecture; S13: Model Training and Optimization: The Adam optimizer and MSE loss function are used to iteratively optimize parameters through the training set, and an early stopping mechanism is used to suppress overfitting. S14: Model Testing and Consolidation: Verify generalization ability through the test set. When MAE ≤ ±3cm, consolidate the model parameters.

[0052] Step S2: Synchronously acquire water level and air pressure time-series data within the gas chamber of the oscillating water column wave energy generator. This is done through a data acquisition unit that simultaneously acquires the water level and air pressure time-series data within the gas chamber. Specifically, Step S2 includes: acquiring raw data using a submersible level gauge and a pressure transmitter at a synchronous sampling frequency of 5~20Hz. The synchronous acquisition control module ensures a timestamp error ≤10ms.

[0053] Step S3: Input the water level time series data and air pressure time series data into the data preprocessing unit. After cleaning, normalization, and time series feature construction, the feature mapping capability is calculated through the loss boundary-mapping capability feature screening module, and effective features are retained to obtain an effective feature set. Step S3 specifically includes: S31: Data Cleaning: 3 The criteria are used to remove outlier data, linear interpolation to fill in missing data, and moving average filtering to remove noise. S32: Data normalization: Min-Max normalization is mapped to the [0,1] interval; S33: Temporal Feature Construction: Constructing water level, air pressure, and their derived feature sets; S34: LM Feature Selection: Calculates the mapping ability of each feature. ,reserve The effective features are used to form the final input feature set.

[0054] Step S4: Input the effective feature set into the trained AI tide prediction model for inference and prediction, and output the predicted tide value. Step S4 specifically includes: inputting the feature set into the customized TCN benchmark model, and outputting the current tide prediction value and the prediction sequence for a preset future time in real time.

[0055] Step S5: The predicted tide level is denormalized to restore the actual tide level. The actual tide level is compared with the data from the tide gauge station daily. When the average absolute error exceeds ±5cm for three consecutive days, an incremental training set is constructed using the newly added data to adjust the artificial intelligence tide level prediction model and replace the original artificial intelligence tide level prediction model, thus achieving online adaptive updating of the model. Step S5 specifically includes: S51: Model performance monitoring: Daily comparison of the actual physical quantity of tide level after inverse normalization with the baseline data to trigger update conditions; S52: Incremental update training: Fine-tuning model parameters based on new data and replacing the original model.

[0056] Preferably, in some embodiments of the present invention, the method further includes: Step S6: Output and engineering application of the predicted tide level value, wherein the actual physical quantity of the tide level after denormalization is transmitted to the local controller and remote monitoring platform of the oscillating water column wave energy power generation device through the tide level output and application docking unit, and data storage and anomaly alarm are completed simultaneously. Step S6 specifically includes: outputting tide level data to the device controller and remote monitoring platform for operation control and maintenance decisions, and data storage and anomaly alarm are completed simultaneously.

[0057] Example 3, Experimental Comparative Example.

[0058] (I) Experimental Environment and Equipment Parameters 1. Experimental location: An offshore OWC test power station (latitude 30°N, longitude 122°E). The sea area has a semi-diurnal tide type, with an average tidal range of 2.5m, a maximum tidal range of 4.2m, and a significant wave height range of 0.3~5.2m. 2. Data acquisition equipment: ① Air chamber water level sensing module: Submersible level gauge, model: JL-800, measuring range 0~5m, accuracy ±1cm, sampling frequency 10Hz; ② Air chamber pressure sensing module: pressure transmitter, model: PTL501, measurement range 80~120kPa, accuracy 0.1%FS, sampling frequency 10Hz; ③ Synchronous acquisition and control module: STM32H743 microcontroller, 485 bus communication, time synchronization accuracy ≤10ms; ④ Tide level label acquisition equipment: Pressure type tide gauge, model: Solinst Levelogger Edge, measuring range 0~10m, accuracy ±1cm, GPS time synchronization; 3. Model training and inference equipment: Training equipment: CPU Intel Core i7-12700K, GPU NVIDIA RTX 3090, RAM 32GB, Python 3.9, PyTorch 2.0; Inference device: Edge controller, model: NVIDIA Jetson Xavier NX, computing power 21 TOPS, deployed in the local control cabinet of the OWC device.

[0059] (II) Experimental Design and Comparison Scheme To verify the technical advantages of this invention, four sets of comparative experiments were set up. Each set of experiments used the same training set, validation set, and test set. The test set contained 5000 samples of normal sea state and 1000 samples of extreme high-wave sea state. The evaluation indicators included mean absolute error (MAE), root mean square error (RMSE), accuracy of tidal trend prediction (error ≤ 5cm is considered accurate), and single-sample inference time. Simultaneously, an "LM screening module ablation experiment" was added to verify the effectiveness of the core innovation. 1. Comparison with Scheme 1: Traditional submersible water level gauge direct measurement method, without AI processing, directly outputs the water level gauge measurement value as tide level data; 2. Comparison with Scheme 2: Traditional LSTM tide prediction model, with the same input features as this scheme, and model structure: input layer → 2 layers of LSTM hidden layers (128 units each) → fully connected output layer; 3. Comparison with Scheme 3: The original TCN time series prediction model has the same input features as this scheme. The model structure is: input layer → 3 TCN hidden layers (inflation factor d=1,2,4) → fully connected output layer; 4. Comparison with Scheme 4: The scheme of this invention includes LM screening + customized TCN model; 5. Ablation scheme: The scheme of this invention removes the LM screening module, while the rest of the structure is completely the same.

[0060] (III) Experimental Results and Analysis 1. Experimental results under normal sea conditions: Table 1 Results of routine sea state experiments

[0061] 2. Under extreme sea conditions with large waves ( >4m) Experimental results: Table 2 Results of extreme sea state experiments

[0062] 3. Results Analysis: Under normal sea conditions, the MAE of the proposed solution is only ±2.3cm, which is 60.3% lower than the traditional LSTM model and 46.5% lower than the original TCN model; the MAE of the ablation solution is ±3.7cm, proving that the LM screening module can improve the accuracy by 37.8%, highlighting its core innovative value. Under extreme high-wave conditions, the MAE of the proposed solution is still controlled within ±4.2cm, which is 74.5% lower than the traditional LSTM model and 61.1% lower than the original TCN model; the MAE of the ablation solution is ±8.5cm, further verifying the key role of the LM screening module in suppressing wave noise. Advantages of the customized TCN architecture: Spatiotemporal coupling feature extraction solves the nonlinear mapping problem between water level and air pressure, and the piecewise expansion factor strategy solves the problem of capturing long-period tides. The synergy of the two makes the model far superior to existing models in multi-scale signal separation. In terms of inference efficiency, the single-sample inference time of the present invention is ≤5ms, which meets the real-time control requirements of OWC devices, and the inference efficiency is improved by 50% compared with LSTM model; Long-term stability verification: After 6 months of continuous operation, the MAE of the proposed solution was ±3.4cm, which was only 1.1cm higher than the initial state. In contrast, the MAE of the comparative solutions 2 and 3 increased to ±8.7cm and ±7.5cm, respectively, proving the effectiveness of the online adaptive update mechanism.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0064] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An artificial intelligence tide prediction system based on water level and air pressure data in an air chamber, characterized in that, The system, applied to an oscillating water column wave energy generation device, comprises: The data acquisition unit is used to simultaneously acquire the water level time-series data and air pressure time-series data in the air chamber of the device; A data preprocessing unit, connected to the data acquisition unit, is used to preprocess the water level time series data and the air pressure time series data. The preprocessing includes: constructing multiple candidate features, and selecting effective features from the multiple candidate features based on the mapping contribution between each candidate feature and the external tide level label. The artificial intelligence tide prediction model unit, connected to the data preprocessing unit, includes an artificial intelligence tide prediction model that uses a time-series convolutional network customized and improved for the oscillating water column device scenario. The customized and improved time-series convolutional network is used to extract the fluid-structure interaction spatial features of water level and air pressure and the time features of low-frequency tidal trends based on the effective features and complete the spatiotemporal fusion. At the same time, it captures the long-period temporal dependence of tides and outputs the tide prediction value. The customized and improved temporal convolutional network includes: a feature extraction layer, used to perform spatiotemporal coupling processing on the effective features, extracting fluid-structure interaction spatial features of water level and air pressure through spatial dimension channel selection, filtering high-frequency wave noise and retaining low-frequency tidal trend temporal features through temporal dimension gated convolution, and outputting fused features after fusing the fluid-structure interaction spatial features and low-frequency tidal trend temporal features; multiple temporal convolutional hidden layers, connected to the feature extraction layer, used to perform temporal convolution processing on the fused features to capture long-period temporal dependencies of tides and output deep temporal features; and a fully connected output layer, connected to the temporal convolutional hidden layers, used to map the deep temporal features to a single-dimensional external tide level prediction value.

2. The artificial intelligence tide prediction system based on water level and air pressure data in an air chamber according to claim 1, characterized in that, The data acquisition unit includes a chamber water level sensing module, a chamber air pressure sensing module, and a synchronous acquisition control module. The air chamber water level sensing module is an immersion-type liquid level gauge used to collect time-series data of the water level inside the air chamber; The air chamber pressure sensing module is a high-precision pressure transmitter with the same sampling frequency as the air chamber water level sensing module, used to collect time-series air pressure data inside the air chamber. The synchronous acquisition and control module performs synchronous management and control of the submersible level gauge and the high-precision pressure transmitter.

3. The artificial intelligence tide prediction system based on water level and air pressure data in an air chamber according to claim 1, characterized in that, The data preprocessing unit includes a loss boundary-mapping capability feature filtering module, which is used for: Determine the loss boundary value, which is the minimum theoretical loss value that a preset single-feature benchmark regression model can achieve when inputting random noise; For each candidate feature, determine the actual loss value for that candidate feature. The actual loss value is the loss value obtained after training the preset single-feature benchmark regression model when the candidate feature is used as the only input. Based on the loss boundary value and the actual loss value, calculate the mapping capability quantization value of the candidate feature; Candidate features with a mapping capability quantization value greater than zero are retained as valid features.

4. The artificial intelligence tide prediction system based on water level and air pressure data in an air chamber according to claim 3, characterized in that, The formula for calculating the loss boundary value in the loss boundary-mapping capability feature filtering module is as follows: ； In the formula, This represents the boundary value of the loss. This represents the total number of samples in the pre-training dataset; For the first Random noise prediction value for each sample; For the first The actual external tide level label value for each sample; For the first in the dataset One sample.

5. The artificial intelligence tide prediction system based on water level and air pressure data in an air chamber according to claim 1, characterized in that, The data preprocessing unit includes a data cleaning module, a data normalization module, and a time-series feature construction module; The data cleaning module uses a three-standard-deviation criterion to remove outliers, linear interpolation to fill missing data, and a five-step moving average filtering to remove noise. The time-series feature construction module constructs a multi-dimensional feature set including historical values ​​of water level and air pressure, water level change rate and air pressure change rate per unit time, cross-product features, and statistical features.

6. The artificial intelligence tide prediction system based on water level and air pressure data in an air chamber according to claim 1, characterized in that, The feature extraction layer includes: The spatial dimension channel selection unit is used to dynamically adjust the weight of each feature channel according to the fluid-structure interaction characteristics of water level and air pressure, strengthen the feature channels related to tides and suppress the wave noise channel, so as to output the fluid-structure interaction spatial characteristics. The time-dimensional gated convolutional unit is used to filter high-frequency wave noise through the reset gate and retain low-frequency tidal trend features through the update gate, so as to output low-frequency tidal trend time features. The dual-path feature fusion unit is used to fuse the fluid-structure interaction spatial features output by the spatial dimension channel selection unit with the low-frequency tidal trend temporal features output by the temporal dimension gated convolution unit to obtain fused features. The fused features are then processed by batch normalization and activation functions and used as the output of the feature extraction layer.

7. The artificial intelligence tide prediction system based on water level and air pressure data in an air chamber according to claim 1, characterized in that, The multiple temporal convolutional hidden layers are configured using a segmented re-dilation factor strategy; in the segmented re-dilation factor strategy, in at least one pair of adjacent temporal convolutional hidden layers, the dilation factor of the later layer is less than or equal to the dilation factor of the earlier layer.

8. The artificial intelligence tide prediction system based on water level and air pressure data in an air chamber according to claim 1, characterized in that, The system also includes an online adaptive model update unit and a tide level output and application interface unit. The online adaptive model update unit includes: The model performance monitoring module is used to monitor the prediction performance of the artificial intelligence tide level prediction model unit and determine whether the prediction performance is lower than a preset threshold. An online incremental training module is used to perform incremental fine-tuning training on the artificial intelligence tide level prediction model unit based on newly added on-site collected data when the prediction performance is lower than a preset threshold, so as to update the model parameters. The tide level output and application docking unit realizes the output and storage of tide level prediction values ​​and docking with the engineering application of the oscillating water column wave energy power generation device.

9. An artificial intelligence-based tidal level prediction method based on water level and air pressure data in an air chamber, characterized in that, The system applied to any one of claims 1 to 8 includes the following steps: Step S1: Construct a full-scene dataset containing water level and air pressure data of the air chamber of the oscillating water column wave energy power generation device and corresponding tide level labels, complete the offline pre-training of the artificial intelligence tide level prediction model, solidify the model parameters, and obtain the trained artificial intelligence tide level prediction model. Step S2: Synchronously collect water level and air pressure time series data in the air chamber of the oscillating water column wave energy generator. The water level and air pressure time series data inside the air chamber of the oscillating water column wave energy generator are collected synchronously through the data acquisition unit. Step S3: Input the water level time series data and air pressure time series data into the data preprocessing unit, and perform cleaning, normalization and time series feature construction in sequence. Then, calculate the feature mapping capability through the loss boundary-mapping capability feature screening module and retain the effective features to obtain the effective feature set. Step S4: Input the effective feature set into the trained artificial intelligence tide prediction model for inference and prediction, and output the tide prediction value; Step S5: The predicted tide level is denormalized to restore the actual physical quantity of the tide level. The actual physical quantity of the tide level is compared with the data from the tide gauge station every day. When the average absolute error exceeds ±5cm for 3 consecutive days, the artificial intelligence tide level prediction model is adjusted by constructing an incremental training set with the newly added data and the original artificial intelligence tide level prediction model is replaced, so as to realize the online adaptive update of the model.

Images