BiLSTM and adaptive attention mechanism fused offshore wind power prediction method and system

Abstract

The invention discloses an offshore wind power prediction method and system fusing BiLSTM (Bidirectional Long Short Term Memory) and an adaptive attention mechanism. The method comprises a data acquisition and treatment module, a feature processing module, a depth prediction module and a visual interface module, and supports multi-source data access and multi-time-domain power prediction. The method comprises the steps that based on a deep learning model, a comprehensive prediction framework fusing a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM) and a self-attention mechanism is provided, the CNN extracts local features of an input meteorological and power sequence, the BiLSTM captures a bidirectional time dependency relationship, the self-attention mechanism introduces an adaptive adjustment factor, and the self-attention mechanism is used for predicting the local features of the input meteorological and power sequence. The weight is dynamically adjusted according to the wind speed change rate, the wind direction, the barometric gradient and other characteristics, and the recognition capacity of the model for key time periods and abrupt change signals is improved. According to the method, the prediction precision and robustness under the extreme wind condition can be remarkably improved, the power grid dispatching safety and the new energy consumption capacity are improved, and technical support is provided for an offshore wind plant.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation and intelligent power dispatch, and more specifically, relates to an offshore wind power prediction method and system that integrates BiLSTM and adaptive attention mechanism. Background Technology

[0002] With the rapid development of offshore wind power, its installed capacity and grid connection scale are constantly expanding, and the non-stationarity and volatility of wind power are becoming increasingly prominent. Since offshore wind speed is affected by a variety of meteorological and oceanographic factors, such as pressure gradient, humidity changes, sea surface temperature difference, wind shear, and typhoon peripheral disturbances, wind power prediction has become a key technical challenge in grid security dispatch and renewable energy consumption.

[0003] Existing wind power forecasting methods mainly include numerical meteorological models based on physical mechanisms and time series models based on statistical learning. The former relies on a large number of meteorological parameters and complex numerical solutions, resulting in high computational costs and insufficient response to local offshore characteristics; the latter, such as ARIMA, support vector regression (SVR), or one-way LSTM, can capture time series relationships to some extent, but their accuracy is limited when dealing with strongly nonlinear, non-stationary, and multivariate coupled characteristics, and they cannot effectively reflect the impact of sudden wind changes on power output.

[0004] In offshore wind farms, power output often exhibits significant randomness and nonlinearity. Affected by sudden wind speed changes, thermal and humidity interactions, and typhoon edge disturbances, traditional models are prone to prediction lag and accuracy drift, making it difficult to meet the time precision requirements of actual dispatching. Therefore, there is an urgent need for a prediction system capable of integrating multi-source meteorological information and possessing temporal memory and adaptive learning capabilities to achieve high-precision and robust offshore wind power prediction. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for offshore wind power prediction that integrates BiLSTM and an adaptive attention mechanism. This system can comprehensively utilize multi-source meteorological data and wind turbine operation data, and through collaborative modeling using a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an adaptive attention mechanism, it can achieve high-precision prediction and dynamic response of wind power under complex sea conditions.

[0006] To achieve the above objectives, the present invention provides an offshore wind power prediction system integrating BiLSTM and an adaptive attention mechanism, comprising: S100: Collects multi-source meteorological and power data, including characteristic variables such as wind speed, wind direction, temperature, humidity, air pressure, cloud cover, radiation, precipitation and historical power, and performs data cleaning and unification through outlier removal, missing value imputation and normalization methods. S200: Extracts spatial correlations and local patterns among input features, and extracts representative features through convolutional layers, activation layers and pooling layers to provide compressed and efficient feature inputs for subsequent temporal modeling; S300: Establish a bidirectional dependency model for time series, dynamically allocate feature weights by combining a self-attention mechanism, introduce a learnable dynamic offset adjustment factor, and adjust the attention distribution in real time according to external conditions such as wind speed change rate, wind direction and pressure gradient, thereby enhancing the model's ability to respond to sudden changes in wind conditions and seasonal fluctuations. S400: Outputs prediction results, error analysis, uncertainty assessment and model interpretability, provides comparison between predicted power and measured power, confidence interval visualization and attention weight heatmap in multiple dimensions, and supports real-time data interaction with scheduling system and SCADA system.

[0007] Further, in step S100, the data acquisition and processing includes: S101: Detect and correct outliers in the observation data caused by sensor failure, communication interruption or extreme noise, using statistical testing methods or smoothing replacement methods based on time-series context to ensure the rationality of data distribution. S102: For missing data due to equipment failure or harsh environment during data acquisition, interpolation methods are used to fill the gaps. Linear interpolation is used for short time intervals of missing data, and spline interpolation is used for long time or nonlinear missing data to ensure the continuity and integrity of time series data. S103: The min-max normalization method is adopted to linearly map all input features to the [0,1] interval, thereby eliminating the dimensional differences between different physical quantities and ensuring the balanced weight distribution of each variable during the model training process.

[0008] Furthermore, the normalization process in the data acquisition and governance module is calculated using this formula: , In the formula, These are the normalized feature variable values; These are the original feature variable values; The maximum value among the original feature variable values. It is the minimum value among the original feature variable values.

[0009] Further, in step S200, extracting the spatial correlation and local patterns among the input features includes: S201: By performing weighted summation on the input multidimensional feature matrix through a sliding convolution kernel, the local correlation and pattern features between different meteorological variables are captured. S202: By using nonlinear functions (such as ReLU) to introduce nonlinear expressive capabilities, the model can adapt to complex coupling relationships such as wind speed, air pressure and temperature. S203: Perform dimensionality reduction on the convolution results to retain significant features while reducing computational complexity; S204: Map and combine the extracted multidimensional local features to generate a high-level abstract feature vector suitable for time series prediction.

[0010] Furthermore, in step S300, the bidirectional dependency model of the time series includes a bidirectional long short-term memory network, which consists of two independent LSTM networks: one for processing the input sequence in the forward direction and the other for processing the sequence in the reverse direction, and the hidden states of the two are fused at the output.

[0011] Furthermore, the information transmission process of the bidirectional long short-term memory network is controlled collaboratively by three gating units: Forgotten Gate: , In the formula, In time step The output of the forget gate, and These represent the weight matrix and bias term of the forget gate, respectively. The Sigmoid activation function maps the input to the interval [0,1] to control the degree to which information is retained or forgotten. Input Gate: , , , In the formula, In time step The output state of the input gate is used to determine the current input. Cell state The extent of the impact of the update and These are the weight matrix and bias term of the input gate, respectively. and These are the generated post-selection states. The weight matrix and bias terms are given, and tanh is the hyperbolic tangent activation function used for nonlinear transformation of candidate states. This is for element-wise multiplication operations, used to implement weighted control of different feature dimensions in the gating mechanism; Output gate: , In the formula, For time steps The output state of the output gate. and These are the weight matrix and bias term corresponding to the output gate, respectively, used to adjust the amount of information transmitted from the current state to the external hidden layer, thereby achieving dynamic control of the model output.

[0012] Furthermore, the adaptive attention mechanism can adaptively evaluate the importance of each input segment in a specific task and dynamically adjust the weight distribution of the information flow accordingly, specifically including: Weight calculation: Calculate the similarity or relevance score between the query vector and a set of key vectors to measure the model's attention to different input locations. , Normalization: The Softmax function is used to normalize all scores, converting them into attention weights in the form of a probability distribution to ensure that the sum of the weights is 1. , Weighted aggregation: Based on the normalized weights, the corresponding value vectors are summed in a weighted manner to obtain a weighted context vector, which serves as the final attention output. , In the formula, To score attention, This represents the normalized attention weights. For weighted context vectors, For the input feature vector, The bias term for the global offset. This is the weight matrix used to project the input features into the attention space.

[0013] Furthermore, the adaptive attention mechanism adds a dynamic offset term. : , In the formula, For time steps The wind speed at the location is affected by learnable weight parameters. Adjustments should be made. Wind direction information is represented by a weight matrix. Projected onto offset space, it is used to characterize the impact of wind direction changes on system dynamics. The wind speed change rate is used to characterize abrupt changes in wind speed. The model's response sensitivity to this characteristic is determined by a coefficient. control, and These are learnable parameters used to capture periodic patterns in time series, such as diurnal variations or seasonal fluctuations.

[0014] Furthermore, step S400 includes: The system displays ultra-short-term / short-term / day-ahead forecasts in layers using a unified timeline, supports the overlay of "predicted power - measured power" curves, residual curves and error distribution maps, and provides scrolling window viewing and time resolution switching. Output quantile prediction band / confidence interval, and simultaneously display reliability indicators such as coverage and calibration curve, so that the scheduling side can use risk-aware threshold configuration in standby and power limiting strategies; The system aggregates and displays RMSE, MAE, MAPE, and R² indicators by time period / unit / wind field, and supports comparison between model versions, before and after feature switching, and different sea state classifications. It also allows users to click on anomalies to trace back to the original feature fragments and the intermediate output of the model for backtracking and location.

[0015] According to a second aspect of the present invention, an offshore wind power forecasting system integrating BiLSTM and an adaptive attention mechanism is provided, comprising: The data acquisition and processing module is used to collect multi-source meteorological and power data, including characteristic variables such as wind speed, wind direction, temperature, humidity, air pressure, cloud cover, radiation, precipitation and historical power, and to clean and unify the data through outlier removal, missing value imputation and normalization methods. The feature processing module is used to extract spatial correlations and local patterns among input features. It extracts representative features through convolutional layers, activation layers and pooling layers, providing compressed and efficient feature inputs for subsequent temporal modeling. The deep prediction module is used to build a bidirectional dependency model for time series. It dynamically allocates feature weights by combining a self-attention mechanism and introduces a learnable dynamic offset adjustment factor. It adjusts the attention distribution in real time according to external conditions such as wind speed change rate, wind direction and air pressure gradient, thereby enhancing the model's ability to respond to sudden changes in wind conditions and seasonal fluctuations. The visualization interface module is used to output prediction results, error analysis, uncertainty assessment and model interpretability. It provides multi-dimensional information such as comparison of predicted power and measured power, visualization of confidence intervals and attention weight heatmaps, and supports real-time data interaction with scheduling systems and SCADA systems.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The system of this invention can access multi-source meteorological and historical power data of wind turbines in real time, and realize efficient data flow and prediction output through cloud-edge collaborative deployment. The modular architecture is adaptable to different wind farms and supports rapid migration and online model updates. Among them, the visualization interface module uses curve overlay, heat map and quantile prediction band to realize the interpretability and uncertainty expression of prediction results, help operation and maintenance and dispatch personnel to grasp power trends and avoid risks, thereby improving the safety of power grid dispatch and the ability of new energy to be absorbed.

[0017] 2. The method of this invention effectively improves the ability to model the spatiotemporal coupling features of offshore wind power sequences by integrating CNN, BiLSTM and Attention mechanisms. CNN can accurately capture the local spatial correlation between wind speed and multiple meteorological variables such as air pressure, humidity and radiation. BiLSTM can bidirectionally mine the long-term and short-term time dependencies of the power sequence. The combination of the two and the Attention mechanism to focus on key influencing periods significantly improves the prediction accuracy of traditional models under complex meteorological conditions such as sudden wind speed changes, strong wind shear, and heat-humidity interaction, and reduces prediction bias caused by the coupling of multiple factors.

[0018] 3. The method of the present invention introduces a dynamic offset adjustment factor (δ). t This factor can be calculated in real time based on core meteorological characteristics such as wind speed, wind direction, and wind speed change rate, and can be learned through weights. W v 、W θ To adapt to the degree of influence of different meteorological elements, by using coefficients τ 1. Enhances the sensitivity to sudden changes in wind speed, thereby adaptively adjusting the attention distribution; in sudden scenarios such as rapid increases or decreases in wind speed and sharp changes in wind direction, it can quickly focus attention weights on key influencing features, effectively avoiding the lag of traditional fixed-weight mechanisms; simultaneously, this factor can also... τ 2. τ 3. Capture periodic features such as diurnal rhythms and seasonal alternations to achieve sensitive capture of power fluctuations in multiple time domains, including ultra-short-term, short-term, and day-ahead periods. Simultaneously correct prediction deviations caused by dynamic changes in meteorological conditions to achieve error self-correction. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the offshore wind power prediction method integrating BiLSTM and adaptive attention mechanism according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the feature processing module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the long short-term memory network in the deep prediction module of this invention. Figure 4 This is a schematic diagram of the bidirectional long short-term memory network in the deep prediction module of this invention. Figure 5 This is a schematic diagram of the attention mechanism in the depth prediction module of this invention. Figure 6 This is a schematic diagram of the overall process of an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] like Figure 1-6 As shown, this invention provides an offshore wind power forecasting system integrating BiLSTM and adaptive attention mechanisms. The system includes a data acquisition and governance module, a feature processing module (CNN), a deep prediction module (BiLSTM + Attention), and a visualization interface module. The data acquisition and governance module collects multi-source meteorological and power data, including feature variables such as wind speed, wind direction, temperature, humidity, air pressure, cloud cover, radiation, precipitation, and historical power. It performs data cleaning and unification through outlier removal, missing value imputation, and normalization methods. The feature processing module uses a convolutional neural network (CNN) to extract spatial correlations and local patterns between input features. It extracts representative features through convolutional layers, activation layers, and pooling layers, providing compressed and efficient feature inputs for subsequent time-series modeling. The deep prediction module uses a bidirectional long short-term memory network (BiLSTM) to achieve bidirectional dependency modeling of time series data. It dynamically allocates feature weights using a self-attention mechanism and introduces a learnable dynamic offset adjustment factor (BiLSTM + Attention). The model adjusts its attention distribution in real time based on external conditions such as wind speed change rate, wind direction, and pressure gradient, thereby enhancing its responsiveness to sudden wind changes and seasonal fluctuations. The visualization interface module is used for displaying prediction results, error analysis, uncertainty assessment, and model interpretability output. It provides multi-dimensional displays such as comparisons between predicted and measured power, confidence interval visualization, and attention weight heatmaps, and supports real-time data interaction with scheduling and SCADA systems.

[0022] The data acquisition and processing module is used to collect, clean, and standardize multi-source observation data related to the operation and environment of offshore wind farms. This data includes, but is not limited to, the following characteristic variables: atmospheric pressure, relative humidity, cloud cover, wind speed (at 10m and 100m altitude), wind direction (at 10m and 100m altitude), temperature, radiation intensity, and precipitation. These characteristic variables directly or indirectly affect the generation of offshore wind power. For example, atmospheric pressure is related to air density, wind speed is directly linked to wind energy capture efficiency, temperature, humidity, and cloud cover reflect local weather evolution, radiation intensity is related to Earth-sea heat exchange, and precipitation is often accompanied by frontal or convective activity leading to sudden changes in wind speed.

[0023] The data acquisition and governance module specifically includes three steps: (1) Outlier removal: Outliers in the observation data caused by sensor failure, communication interruption or extreme noise are detected and corrected by using statistical test methods or smoothing replacement methods based on time context to ensure the rationality of data distribution.

[0024] (2) Missing value imputation: For missing values ​​caused by equipment failure or harsh environment during data acquisition, interpolation method is used to fill them. Linear interpolation is used for short time interval missing values, and spline interpolation is used for long time or non-linear missing values ​​to ensure the continuity and integrity of time series data.

[0025] (3) Normalization: The min-max normalization method is adopted to linearly map all input features to the [0,1] interval, thereby eliminating the dimensional differences between different physical quantities and ensuring the balanced weight distribution of each variable during the model training process.

[0026] The normalization process uses this formula for calculation: , In the formula, These are the normalized feature variable values; These are the original feature variable values; The maximum value among the original feature variable values. It is the minimum value among the original feature variable values.

[0027] like Figure 1As shown, the feature processing module is used to perform deep feature extraction on multi-source meteorological and operational features after data collection and normalization, providing more representative input for subsequent time series modeling. The core of the module uses a Convolutional Neural Network (CNN), whose basic units include convolutional layers, activation layers, pooling layers, and fully connected layers. In this module, the convolutional layers perform weighted summation on the input multi-dimensional feature matrix using sliding convolution kernels, thereby capturing the local correlations and pattern features between different meteorological variables; the activation layers use nonlinear functions (such as ReLU) to introduce nonlinear expressive power, enabling the model to adapt to complex coupling relationships such as wind speed, air pressure, and temperature; the pooling layers perform dimensionality reduction on the convolution results, preserving significant features while reducing computational complexity; and the fully connected layers map and combine the extracted multi-dimensional local features to generate high-level abstract feature vectors suitable for time series prediction. The feature processing module, utilizing CNN, can automatically extract features from the input data and compress feature dimensions, thereby enhancing the model's generalization ability. In power load forecasting tasks, convolutional layers use convolutional kernels to extract local patterns, while pooling layers highlight important features and reduce computational complexity by reducing dimensionality.

[0028] , In the formula, For the l-th layer Each feature mapping, For activation function, For convolution operations, To convolve this weight matrix, This represents the P-th input feature map from the first layer. This is a bias term.

[0029] , In the formula, For pooling output, the function This indicates a max pooling operation. Input feature vector, This represents the size of the pooling window.

[0030] , In the formula, The feature vector output by the network. and These are represented as the weight matrix and the bias term, respectively. For input variables.

[0031] , In the formula, The total number of samples in the dataset. For the loss function, and These are the actual output and the predicted output, respectively. This is the set of model parameters obtained after optimization.

[0032] The deep prediction module specifically includes a deep prediction model based on a bidirectional long short-term memory network (BiLSTM) and an adaptive attention mechanism. This module is used to dynamically model the temporal features extracted by convolution to capture the time dependence and non-stationarity of offshore wind power sequences under complex ocean and atmospheric conditions.

[0033] like Figure 3 and Figure 4 As shown, the core of the deep prediction module employs a bidirectional Long Short-Term Memory (BiLSTM) network. This structure consists of two independent LSTM networks: one for processing the input sequence in the forward direction (from start to finish), and the other for processing the sequence in the reverse direction (from finish to start). The hidden states of both networks are fused at the output. Through this bidirectional structure, BiLSTM can not only identify the impact of historical information on current power changes but also comprehensively utilize the potential feedback relationships between future timeframes and preceding states, achieving bidirectional dependency modeling in the time dimension. This mechanism is particularly suitable for the characteristics commonly found in offshore wind farm power sequences, such as hysteresis response, periodic disturbances, and sudden wind speed changes.

[0034] The information transfer process in the bidirectional long short-term memory network is controlled collaboratively by three gating units: (1) Gate of Oblivion: , In the formula, In time step The output of the forget gate, and These represent the weight matrix and bias term of the forget gate, respectively. The Sigmoid activation function maps the input to the interval [0,1] to control the degree to which information is retained or forgotten.

[0035] (2) Input gate: , , , In the formula, In time step The output state of the input gate is used to determine the current input. Cell state The extent of the impact of the update and These are the weight matrix and bias term of the input gate, respectively. and These are the generated post-selection states. The weight matrix and bias terms are given, and tanh is the hyperbolic tangent activation function used for nonlinear transformation of candidate states. This is for element-wise multiplication operations, used to implement weighted control of different feature dimensions in gating mechanisms.

[0036] (3) Output gate: , In the formula, For time steps The output state of the output gate. and These are the weight matrix and bias term corresponding to the output gate, respectively, used to adjust the amount of information transmitted from the current state to the external hidden layer, thereby achieving dynamic control of the model output.

[0037] like Figure 5 As shown, the core idea of ​​the attention mechanism is to characterize the degree of attention the model assigns to different parts of the input data through learnable weight allocation. This mechanism can adaptively evaluate the importance of each input segment in a specific task and dynamically adjust the weight distribution of the information flow accordingly. Specifically, it includes three steps: (1) Weight calculation Calculate the similarity or relevance score between the query vector and a set of key vectors to measure the model's attention at different input locations: , (2) Normalization The softmax function is used to normalize all scores, transforming them into attention weights in the form of a probability distribution to ensure that the sum of the weights is 1. , (3) Weighted aggregation Based on the normalized weights, the corresponding value vectors are summed in a weighted manner to obtain a weighted context vector, which serves as the final attention output. , In the formula, To score attention, This represents the normalized attention weights. For weighted context vectors, For the input feature vector, The bias term for the global offset. This is the weight matrix used to project the input features into the attention space.

[0038] The attention mechanism adds a dynamic offset term ( Dynamic Offset Term): , In the formula, For time steps The wind speed at the location is affected by learnable weight parameters. Adjustments should be made. Wind direction information is represented by a weight matrix. Projected onto offset space, it is used to characterize the impact of wind direction changes on system dynamics. represents the rate of change of wind speed, used to characterize abrupt changes in wind speed. The model's sensitivity to this characteristic is determined by a coefficient. control, and These are learnable parameters used to capture periodic patterns in time series, such as diurnal variations or seasonal fluctuations. This is achieved by introducing... The model can dynamically adjust its attention distribution based on the intensity of sudden wind speed changes, meteorological disturbances, or air-sea coupling events, thereby maintaining prediction accuracy and stability under complex sea conditions such as high wind shear, the periphery of typhoons, or local turbulence. Furthermore, this mechanism can automatically capture diurnal and seasonal variation patterns in scenarios with significant periodic characteristics, improving the model's cross-temporal generalization ability.

[0039] The visualization interface module is used to present, interact with, and connect to external services the output results and related intermediate features of the deep prediction module, supporting operational shifts, scheduling decisions, and model maintenance. Its functions include, but are not limited to: (1) Display of results in multiple time domains: It displays ultra-short-term / short-term / day-ahead forecasts in layers using a unified time axis, supports the overlay of "predicted power - measured power" curves, residual curves, and error distribution maps; and provides a scrolling window for viewing and switching of time resolution, facilitating local magnification analysis when wind conditions change abruptly.

[0040] (2) Expression of uncertainty and reliability: Output quantile prediction band / confidence interval, and simultaneously display reliability indicators such as coverage and calibration curve, so that the scheduling side can use risk-aware threshold configuration in standby and power limiting strategies.

[0041] (3) Evaluation indicator dashboard: The system aggregates and displays indicators such as RMSE, MAE, MAPE, and R² by time period / unit / wind field, and supports comparison (between model versions, before and after feature switching, and different sea state classifications) and backtracking (click on the anomaly point to backtrack the original feature fragment and the intermediate output of the model).

[0042] like Figure 6 As shown, this embodiment of the invention provides a method for predicting offshore wind power by integrating a bidirectional long short-term memory network and an adaptive attention mechanism. The offshore wind power prediction system, which integrates BiLSTM and an adaptive attention mechanism, includes the following steps: S100: Collects meteorological and wind power data such as wind speed, wind direction, temperature, and air pressure, and eliminates dimensional differences by removing outliers, imputing missing values, and normalizing the data. S200: Convolutional layers capture local correlations of meteorological variables, activation layers adapt to nonlinear relationships, pooling layers reduce dimensionality, and fully connected layers generate high-level abstract feature vectors. S300: Build a BiLSTM network, regulate information transmission through forget gate, input gate, and output gate, fuse bidirectional hidden states, and capture the bidirectional time dependence of power sequence; S400: Calculate attention weights, normalize them using Softmax, and then aggregate them to generate a context vector, introducing a dynamic offset factor δ. t The weights are dynamically adjusted based on wind speed, wind direction, and other factors to adapt to sudden changes in wind conditions.

[0043] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A marine wind power forecasting system integrating BiLSTM and adaptive attention mechanism, characterized in that, include: S100: Collects multi-source meteorological and power data, including characteristic variables such as wind speed, wind direction, temperature, humidity, air pressure, cloud cover, radiation, precipitation and historical power, and performs data cleaning and unification through outlier removal, missing value imputation and normalization methods. S200: Extracts spatial correlations and local patterns among input features, and extracts representative features through convolutional layers, activation layers and pooling layers to provide compressed and efficient feature inputs for subsequent temporal modeling; S300: Establish a bidirectional dependency model for time series, dynamically allocate feature weights by combining a self-attention mechanism, introduce a learnable dynamic offset adjustment factor, and adjust the attention distribution in real time according to external conditions such as wind speed change rate, wind direction and pressure gradient, thereby enhancing the model's ability to respond to sudden changes in wind conditions and seasonal fluctuations. S400: Outputs prediction results, error analysis, uncertainty assessment and model interpretability, provides comparison between predicted power and measured power, confidence interval visualization and attention weight heatmap in multiple dimensions, and supports real-time data interaction with scheduling system and SCADA system.

2. The offshore wind power forecasting system integrating BiLSTM and adaptive attention mechanism according to claim 1, characterized in that, In step S100, the data acquisition and processing includes: S101: Detect and correct outliers in the observation data caused by sensor failure, communication interruption or extreme noise, using statistical testing methods or smoothing replacement methods based on time-series context to ensure the rationality of data distribution. S102: For missing data due to equipment failure or harsh environment during data acquisition, interpolation methods are used to fill in the gaps. Linear interpolation is used for short time interval gaps, and spline interpolation is used for long time or nonlinear gaps to ensure the continuity and integrity of time series data. S103: The min-max normalization method is adopted to linearly map all input features to the [0,1] interval, thereby eliminating the dimensional differences between different physical quantities and ensuring the balanced weight distribution of each variable during the model training process.

3. The offshore wind power forecasting system integrating BiLSTM and adaptive attention mechanism according to claim 2, characterized in that, The normalization process in the data acquisition and processing module is calculated using this formula: ， In the formula, These are the normalized feature variable values; These are the original feature variable values; The maximum value among the original feature variable values. It is the minimum value among the original feature variable values.

4. The offshore wind power forecasting system integrating BiLSTM and adaptive attention mechanism according to any one of claims 1-3, characterized in that, In step S200, extracting the spatial correlation and local patterns among the input features includes: S201: By performing weighted summation on the input multidimensional feature matrix through a sliding convolution kernel, the local correlation and pattern features between different meteorological variables are captured. S202: By using nonlinear functions (such as ReLU) to introduce nonlinear expressive capabilities, the model can adapt to complex coupling relationships such as wind speed, air pressure and temperature. S203: Perform dimensionality reduction on the convolution results to retain significant features while reducing computational complexity; S204: Map and combine the extracted multidimensional local features to generate a high-level abstract feature vector suitable for time series prediction.

5. The offshore wind power forecasting system integrating BiLSTM and adaptive attention mechanism according to any one of claims 1-3, characterized in that, In step S300, the bidirectional dependency model of the time series includes a bidirectional long short-term memory network, which consists of two independent LSTM networks: one for processing the input sequence in the forward direction and the other for processing the sequence in the reverse direction, and the hidden states of the two are fused at the output.

6. The offshore wind power forecasting system integrating BiLSTM and adaptive attention mechanism according to claim 5, characterized in that, The information transmission process of the bidirectional long short-term memory network is controlled collaboratively by three gating units: Forgotten Gate: ， In the formula, In time step The output of the forget gate, and These represent the weight matrix and bias term of the forget gate, respectively. The Sigmoid activation function maps the input to the interval [0,1] to control the degree to which information is retained or forgotten. Input Gate: ， ， ， In the formula, In time step The output state of the input gate determines the current input. Cell state The extent of the impact of the update and These are the weight matrix and bias term of the input gate, respectively. and These are the generated post-selection states. The weight matrix and bias terms are given, and tanh is the hyperbolic tangent activation function used for nonlinear transformation of candidate states. This is for element-wise multiplication operations, used to implement weighted control of different feature dimensions in the gating mechanism; Output gate: ， In the formula, For time step The output state of the output gate. and These are the weight matrix and bias term corresponding to the output gate, respectively, used to adjust the amount of information transmitted from the current state to the external hidden layer, thereby achieving dynamic control of the model output.

7. The offshore wind power forecasting system integrating BiLSTM and adaptive attention mechanism according to claim 6, characterized in that, The adaptive attention mechanism can adaptively evaluate the importance of each input segment in a specific task and dynamically adjust the weight distribution of the information flow accordingly, specifically including: Weight calculation: Calculate the similarity or relevance score between the query vector and a set of key vectors to measure the model's attention to different input locations. ， Normalization: The Softmax function is used to normalize all scores, converting them into attention weights in the form of a probability distribution to ensure that the sum of the weights is 1. ， Weighted aggregation: Based on the normalized weights, the corresponding value vectors are summed in a weighted manner to obtain a weighted context vector, which serves as the final attention output. ， In the formula, To score attention, This represents the normalized attention weights. For weighted context vectors, For the input feature vector, The bias term for the global offset. This is the weight matrix used to project the input features into the attention space.

8. The offshore wind power forecasting system integrating BiLSTM and adaptive attention mechanism according to claim 7, characterized in that, The adaptive attention mechanism adds a dynamic offset term. : ， In the formula, For time step The wind speed at the location is affected by learnable weight parameters. Adjustments should be made. Wind direction information is represented by a weight matrix. Projected onto offset space, it is used to characterize the impact of wind direction changes on system dynamics. The wind speed change rate is used to characterize abrupt changes in wind speed. The model's response sensitivity to this characteristic is determined by a coefficient. control, and These are learnable parameters used to capture periodic patterns in time series, such as diurnal variations or seasonal fluctuations.

9. The offshore wind power forecasting system integrating BiLSTM and adaptive attention mechanism according to claim 1, characterized in that, Step S400 includes: The system displays ultra-short-term / short-term / day-ahead forecasts in layers using a unified timeline, supports the overlay of "predicted power - measured power" curves, residual curves and error distribution maps, and provides scrolling window viewing and time resolution switching. Output quantile prediction band / confidence interval, and simultaneously display reliability indicators such as coverage and calibration curve, so that the scheduling side can use risk-aware threshold configuration in standby and power limiting strategies; The system aggregates and displays RMSE, MAE, MAPE, and R² indicators by time period / unit / wind field, and supports comparison between model versions, before and after feature switching, and different sea state classifications. It also allows users to click on anomalies to trace back to the original feature fragments and the intermediate output of the model for backtracking and location.

10. A marine wind power forecasting system integrating BiLSTM and adaptive attention mechanism, characterized in that, include: The data acquisition and processing module is used to collect multi-source meteorological and power data, including characteristic variables such as wind speed, wind direction, temperature, humidity, air pressure, cloud cover, radiation, precipitation and historical power, and to clean and unify the data through outlier removal, missing value imputation and normalization methods. The feature processing module is used to extract spatial correlations and local patterns among input features. It extracts representative features through convolutional layers, activation layers and pooling layers, providing compressed and efficient feature inputs for subsequent temporal modeling. The deep prediction module is used to build a bidirectional dependency model for time series. It dynamically allocates feature weights by combining a self-attention mechanism and introduces a learnable dynamic offset adjustment factor. It adjusts the attention distribution in real time according to external conditions such as wind speed change rate, wind direction and air pressure gradient, thereby enhancing the model's ability to respond to sudden changes in wind conditions and seasonal fluctuations. The visualization interface module is used to output prediction results, error analysis, uncertainty assessment and model interpretability. It provides multi-dimensional features such as comparison of predicted power and measured power, visualization of confidence intervals and attention weight heatmaps, and supports real-time data interaction with scheduling systems and SCADA systems.

Images