Short-term wind power prediction system and method considering wind speed regulation and power deduction

Abstract

The invention discloses a short-term wind power prediction system and method considering wind speed regulation and power deduction. The system comprises a serialization module, a discretization module, a full connection layer module, a wind speed segmentation control module and an output layer module. Wherein the output end of the serialization module is connected with the input end of the discretization module, the output end of the discretization module is connected with the input end of the full connection layer module, and the wind speed segmentation control module performs segmentation control on the wind speed by using a wind speed track obtained in the serialization module; the output result of the full connection layer module and the output result of the wind speed segmentation control module are fused in the time dimension and then are connected with the input end of the output layer module. The short-term wind power prediction technology considering wind speed regulation and power evolution is provided to solve the technical problems of inaccurate power grid dispatching and increased operation risk caused by inaccurate wind power prediction result due to insufficient wind power plant time sequence power feature extraction in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of renewable energy power generation and comprehensive consumption, and particularly to a short-term wind power prediction technology considering wind speed regulation and power deduction. Background Art

[0002] High wind energy is an important low-carbon energy source. It has many advantages such as low cost, cleanliness, and renewability, and has the potential to become a sustainable energy supply. However, the strong uncertainty and volatility of wind power output make it difficult to accurately grasp the power generation law of wind farms, threatening the safe and stable operation of power systems. Conducting time series correlation analysis on wind farms and introducing it into short-term power prediction to achieve accurate prediction of wind farm time series power is an effective measure to promote wind power consumption and improve the operation stability of power systems.

[0003] Currently, for short-term wind power prediction for the next 4 to 72 hours, the mainstream methods focus on capturing the time series evolution law of wind power itself, or using the causal relationship between it and main influencing factors such as wind speed. These methods usually rely on historical wind power data and meteorological characteristic data provided by numerical weather prediction (NWP). However, the actual influencing factors are complex and changeable, and the information that can be obtained in advance is limited, resulting in many traditional prediction methods being difficult to cope with the changing environment. Deep learning techniques based on the recurrent neural network (RNN) model, especially the long short-term memory neural network (LSTM) and the gated recurrent unit neural network (GRU), can flexibly capture the time dependence of data at different time lengths and have become the main means of current wind power prediction. However, the effectiveness of this method depends on the consistency and uniformity of the time interval of the input time series data. Since the sampling intervals of different variables may vary, and abnormal events such as wind curtailment, power rationing, or instrument failures often occur in actual operation, resulting in inconsistent data sampling times, challenges may be faced when directly applying the RNN model for prediction. To solve the problem of irregular sampling time series, conventional methods usually perform missing value imputation through statistical laws and time dependence relationships. However, when facing wind power data with a high missing rate, the effects of these imputation methods are often not ideal. Another approach is to develop a prediction model that can directly simulate the continuous time evolution process.

[0004] The prior art document "Attentive Neural Controlled Differential Equations for Time-series Classification and Forecasting" discloses a time series classification and forecasting method based on Neural Controlled Differential Equations (NCDE) to solve the problem of irregularly sampled time series. This method continuousizes the discrete sampled time series on the time axis, deduces the missing data on the continuous time axis through numerical integration calculation, and combines an attention mechanism to improve the ability to capture key information, thereby enhancing the model's processing ability for irregular time series data. However, wind power is closely related to external factors such as wind speed and wind direction, as well as the temporal law of wind power itself. The technologies provided by the above prior art documents cannot simultaneously take into account the influence mechanism of external factors on wind power and the temporal law of wind power itself, resulting in inaccurate wind power prediction results, inaccurate power grid dispatching, and increased operation risks.

[0005] Therefore, the applicant proposes a short-term wind power prediction system that considers wind speed regulation and power deduction, which is a hybrid model that can simultaneously take into account both, hoping to overcome the above technical defects. Summary of the Invention

[0006] The object of the present invention is to provide a short-term wind power prediction technology that considers wind speed regulation and power evolution to solve the technical problem that the prior art is insufficient in extracting the temporal power characteristics of a wind farm, resulting in inaccurate wind power prediction results, inaccurate power grid dispatching, and increased operation risks.

[0007] To solve the above technical problem, the technical solution adopted by the present invention is as follows:

[0008] A short-term wind power prediction system that considers wind speed regulation and power deduction, which includes a continuousization module, a discretization module, a fully connected layer module, a wind speed segmented control module, and an output layer module; wherein, the output end of the continuousization module is connected to the input end of the discretization module, the output end of the discretization module is connected to the input end of the fully connected layer module, and the wind speed segmented control module uses the wind speed trajectory obtained in the continuousization module to perform segmented control of the wind speed; after the output result of the fully connected layer module and the output result of the wind speed segmented control module are fused in the time dimension, they are input to the input end of the output layer module.

[0009] When the system is in use, the following steps are adopted:

[0010] Step 1: Collect the historical power data and external wind speed information data of the target wind farm at the minute level through the continuousization module. The data is divided into the power vector of the target wind farm and the wind speed tensor at discrete time points;

[0011] Step 2: The continuousization module continuously converts the wind speed tensor in Step 1 into a wind speed trajectory X(t) through a cubic spline function and obtains the cubic polynomial wind speed trajectory coefficients;

[0012] Step 3: The continuousization module generates a continuous trajectory z(t) containing the implicit features of wind power in the target time period through operations such as integral calculation and dimension mapping on the cubic polynomial wind speed trajectory coefficients obtained in Step 2;

[0013] Step 4: The discretization module concatenates the continuous trajectory z(t) containing the implicit features of wind power in the target time period obtained in Step 3 with the wind speed trajectory X(t) obtained in Step 2, and inputs the concatenation result into the fully connected layer module to obtain the initial prediction result o(t) of the wind power in the target time period;

[0014] Step 5: Process the wind speed trajectory X(t) obtained in Step 2 through the wind speed segmented control module to obtain a weight vector sequence {w i,t , i = 1, 2, 3, 4} for controlling the initial prediction result o(t) of the wind power in the target time period;

[0015] Step 6: The output layer module integrates the initial prediction result o(t) of the wind power in the target time period obtained in Step 4 according to the weight vector sequence {w i,t , i = 1, 2, 3, 4} obtained in Step 5, inputs the integration result into the output layer module, and obtains the wind power prediction curve in the target time period

[0016] The wind power prediction curve can be obtained from the above steps The obtained wind power prediction curve is the prediction result of the short-term wind power.

[0017] In Step 1, the form of the wind speed tensor is {(x0, t0), (x1, t1),..., (x i , t i ), (x i+1 , t i+1 ),..., (x n , t n )}, where (x i , t i ) represents the wind speed value x i at the i-th time point t i .

[0018] In step 2, in each wind speed tensor interval [x i , x i+1 , the cubic spline function S i (x) can be expressed as a cubic polynomial:

[0019] S i (x) = a i + b i (x - x i ) + c i (x - x i ) 2 + d i (x - x i ) 3

[0020] where a i , b i , c i , d i are the cubic polynomial wind speed trajectory coefficients.

[0021] The initial prediction result o(t) of the wind power in the target time period obtained in step 4 is as follows:

[0022]

[0023] In the formula, o(t) is the preliminary prediction result of the target wind power interval in the effective wind speed interval; represents the first fully connected layer, relu is the rectified linear unit function, W o1 and b o1 are the weight and bias coefficients of the first fully connected layer respectively, is the input data of the first fully connected layer; represents the second fully connected layer, tanh is the hyperbolic tangent function, W o2 and b o2 are the weight and bias coefficients of the second fully connected layer respectively, is the output result of the first fully connected layer, which serves as the input of the second fully connected layer.

[0024] The specific operation process of step 6 is as follows:

[0025]

[0026] where "·" represents numerical multiplication, (w 1,t ·0 + w 2,t ·o t + w 3,t ·1 + w 4,t·0) represents an integration operation, which represents the fusion operation of the output result of the fully connected layer module described in claim 1 and the output result of the wind speed segmented control module in the time dimension; the result of the fusion operation is used as the input for the output layer module; represents the wind power value at time t of the final prediction result of the target wind power interval in the effective wind speed interval; o t is the wind power value at time t of the preliminary prediction result of the target wind power interval in the effective wind speed interval; 0 and 1 are the normalization values when the zero power and rated power are output respectively.

[0027] The continuousization module includes a cubic spline function for continuousizing the discrete wind speed and an NHDE model for obtaining the hidden feature state z(t) of the target time period. When using the NHDE model to obtain the hidden feature state z(t) of the target time period, the following steps are adopted:

[0028] Step 3.1: Considering that the wind power is generally proportional to the cube of the wind speed, and the first-order derivative of the wind speed is a quadratic polynomial with respect to the wind speed, so the wind speed power terms not higher than the second power are spliced into a general wind speed trajectory

[0029]

[0030] In the formula, X(t) represents the wind speed trajectory, that is, the first-order vector of the wind speed, and X 2 (t) represents the second-order vector of the wind speed, represents the splicing operation on the vector;

[0031] Step 3.2: Input the general wind speed trajectory obtained in Step 3.1 into a two-layer fully connected network to extract the differential term h x (t) of the wind power controlled by the wind speed, and its simulation process is as follows:

[0032]

[0033] In the formula, W h1 and b h1 are respectively the weight and bias coefficient of the first-layer fully connected layer; W h2 and b h2 are respectively the weight and bias coefficient of the second-layer fully connected layer; relu and tanh respectively represent the rectified linear unit function and the hyperbolic tangent activation function;

[0034] Step 3.3: Combine the wind speed trajectory X(t) and the hidden feature trajectory z(t) to obtain Then input I(t) into the time evolution mechanism module to extract the time inertia information h(t) of the wind power itself; this module is part of the NHDE model, and its specific process is as follows:

[0035]

[0036] where σ represents the sigmoid activation function; p(t) is the update gate, which maps the input to the interval [0, 1] through the sigmoid activation function; is the candidate hidden state, and the gated update result after differentiating it from the previous hidden state z(t) is used to simulate the rate of change of the hidden state over time;

[0037] Step 3.4: Integrate the differential term h x (t) obtained in Step 3.2 and the time inertia information h(t) obtained in Step 3.3 with respect to the wind speed trajectory X(t) and the time term t respectively to simulate the dynamic evolution law of wind power and obtain the hidden feature state z(t) in the target time period:

[0038]

[0039] In the formula, z(t1) represents the hidden feature of the target wind power at time t1, z(t0) is the value at the initial time, t0 and t1 are two adjacent times in the target time period, and the hidden feature states at all adjacent times in the obtained target time period are concatenated in chronological order through the above formula to form z(t).

[0040] In Step 4, the discretization module is used for discretization and concatenation operations. The discretization operation is to discretize the wind speed trajectory X(t) obtained in Step 2 and the continuous trajectory z(t) of the wind power hidden feature obtained in Step 3 in the form of the wind speed tensor {(x0,t0),(x1,t1),...,(x i ,t i ),(x i+1 ,t i+1 ),...,(x n ,t n )}, and the concatenation function is to concatenate the discretized wind speed trajectory and the continuous trajectory z(t) of the discretized wind power hidden feature in the time dimension.

[0041] The fully connected layer module mentioned in Step 4 consists of two fully connected layers.

[0042] The wind speed segmented control module mentioned in Step 5 consists of Figure 1 the multi-layer perceptron model included in, and the steps of using the multi-layer perceptron model are as follows:

[0043] Step 5.1: Divide the input wind speed into four intervals and set the output power rule, requiring that the wind power time series evolution law learned by the NHDE model is only valid when v in < v t < vN is effective within the wind speed range, while at v t < v in or v t > v out , the wind power output is fixed at 0. At v N < v t < v out , it is fixed at the rated output power P N ; where v t and P t represent the input wind speed and the output wind power at time t respectively; f(v t ) is a polynomial function with the input wind speed as the independent variable; v in , v N , v out are the cut-in wind speed, rated wind speed and cut-out wind speed of the wind turbine respectively;

[0044] Step 5.2: Output the predicted wind power weights {w i,t , i = 1, 2, 3, 4} of the model in the four wind speed intervals during the target time period through a multi-layer perceptron with a softmax function in the output layer.

[0045] In step 6, the output layer module consists of two fully connected layers.

[0046] Compared with the prior art, the present invention has the following technical effects:

[0047] 1) Introduce the neural differential equation technology into the short-term wind power prediction, establish the adjacent integral deduction relationship through time continuousization, and provide a new solution for the wind power time series prediction in the scenarios where the numerical samples or there are missing values between the hidden features of the wind power at irregular sampling times;

[0048] 2) Incorporate the time evolution mechanism into the NCDE model to form a hybrid differential neural network (Neural Hybrid Differential Equation, NHDE) model, enabling the NHDE prediction model to simultaneously explore the time series evolution law of the wind power itself and the dynamic non-linear control relationship between the wind speed and the wind power;

[0049] 3) Dynamically divide the wind speed control interval, and introduce a differential segmented control strategy of wind speed-power for different intervals to form a segmented control hybrid differential neural network (Neural Piecewise Hybrid Differential Equation, NPHDE) model, making the adaptability to the dynamic change of the wind speed and the prediction result more reasonable.

[0050] 4) The present invention performs well on the actual wind power generation data set and also has better prediction accuracy compared with other prediction models. Description of the Drawings

[0051] The present invention will be further described below in conjunction with the drawings and embodiments:

[0052] Figure 1 It is a schematic structural diagram of a short-term wind power prediction system considering wind speed regulation and power deduction;

[0053] Figure 2 It is a comparison chart of wind power prediction curves of each model under the condition that the input wind speed has a 10% missing rate;

[0054] Figure 3 It is a comparison chart of wind power prediction curves of each model under the condition that the input wind speed has a 30% missing rate;

[0055] Figure 4 It is a statistical chart of ablation experiment errors;

[0056] Figure 5 It is a comparison chart of ablation model prediction curves;

[0057] Figure 6 It is a graph of error changes corresponding to different numbers of hidden layer neurons;

[0058] Figure 7 It is a graph of error changes corresponding to different learning rates;

[0059] Figure 8 It is a graph of error changes corresponding to different Euler integration step sizes. Detailed Embodiment

[0060] A short-term wind power prediction system considering wind speed regulation and power deduction includes a continuousization module 1, a discretization module 2, a fully connected layer module 3, a wind speed segmented control module 4, and an output layer module 5; wherein, the output end of the continuousization module 1 is connected to the input end of the discretization module 2, the output end of the discretization module 2 is connected to the input end of the fully connected layer module 3, and the wind speed segmented control module 4 uses the wind speed trajectory obtained in the continuousization module 1 to perform segmented control of the wind speed; after the output results of the fully connected layer module 3 and the output results of the wind speed segmented control module 4 are fused in the time dimension, they are input to the input end of the output layer module 5.

[0061] When the system is in use, the following steps are adopted:

[0062] Step 1: The continuousization module 1 collects the minute-level historical power data and external wind speed information data of the target wind farm, and the data is divided into the power vector of the target wind farm and the wind speed tensor at discrete time points;

[0063] Step 2: The continuity module 1 uses the cubic spline function to continuousize the wind speed tensor in Step 1 into the wind speed trajectory X(t), and obtains the cubic polynomial wind speed trajectory coefficients;

[0064] Step 3: The continuity module 1 performs operations such as integral calculation and dimension mapping on the cubic polynomial wind speed trajectory coefficients obtained in Step 2 to generate the continuous trajectory z(t) containing the hidden features of wind power in the target time period;

[0065] Step 4: The discretization module 2 concatenates the continuous trajectory z(t) containing the hidden features of wind power in the target time period obtained in Step 3 with the wind speed trajectory X(t) obtained in Step 2, and inputs the concatenation result into the fully connected layer module 3 to obtain the initial prediction result o(t) of the wind power in the target time period;

[0066] Step 5: The wind speed segmented control module 4 processes the wind speed trajectory X(t) obtained in Step 2 to obtain the weight vector sequence {w i,t , i = 1, 2, 3, 4} for controlling the initial prediction result o(t) of the wind power in the target time period;

[0067] Step 6: The output layer module 5 integrates the initial prediction result o(t) of the wind power in the target time period obtained in Step 4 according to the weight vector sequence {w i,t , i = 1, 2, 3, 4}, inputs the integration result into the output layer module 5, and obtains the wind power prediction curve in the target time period

[0068] The wind power prediction curve can be obtained from the above steps The obtained wind power prediction curve is the prediction result of the short-term wind power.

[0069] In Step 1, the form of the wind speed tensor is {(x0, t0), (x1, t1),..., (x i , t i ), (x i+1 , t i+1 ),..., (x n , t n )}, where (x i , t i ) represents the wind speed value x i at the i-th time point t i .

[0070] In Step 2, on each wind speed tensor interval [x i , x i+1 , the cubic spline function S i (x) can be expressed as a cubic polynomial:

[0071] S i (x) = a i + b i (x - x i ) + c i (x - x i ) 2 + d i (x - x i ) 3

[0072] where a i , b i , c i , d i are the coefficients of the cubic polynomial wind speed trajectory.

[0073] The initial prediction result o(t) of the wind power in the target time period obtained in step 4 is as follows:

[0074]

[0075] In the formula, o(t) is the preliminary prediction result of the target wind power interval in the effective wind speed interval; represents the first fully connected layer, relu is the rectified linear unit function, W o1 and b o1 are the weight and bias coefficients of the first fully connected layer respectively, is the input data of the first fully connected layer; represents the second fully connected layer, tanh is the hyperbolic tangent function, W o2 and b o2 are the weight and bias coefficients of the second fully connected layer respectively, is the output result of the first fully connected layer, which serves as the input of the second fully connected layer.

[0076] The specific operation process of step 6 is as follows:

[0077]

[0078] where "·" represents numerical multiplication, (w 1,t ·0 + w 2,t ·o t + w 3,t ·1 + w 4,t ·0) represents the integration operation, and this integration operation represents the fusion operation of the output result of the fully connected layer module 3 and the output result of the wind speed segmented control module 4 in the time dimension; the result of the fusion operation is used as the input for the output layer module 5; represents the wind power value at time t of the final prediction result of the target wind power interval in the effective wind speed interval; ot is the wind power value at time t for the preliminary prediction result of the target wind power interval within the effective wind speed interval; 0 and 1 are the normalized values at zero power and rated power output respectively.

[0079] The continuous module 1 includes a cubic spline function for continuousizing the discrete wind speed and an NHDE model for obtaining the hidden feature state z(t) in the target time period. When using the NHDE model to obtain the hidden feature state z(t) in the target time period, the following steps are adopted:

[0080] Step 3.1: Considering that the wind power is generally proportional to the cube of the wind speed and the first-order differential of the wind speed is a quadratic polynomial about the wind speed, the pan wind speed trajectory is spliced with the wind speed power terms not higher than the second order.

[0081]

[0082] where X(t) represents the wind speed trajectory, that is, the first-order vector of the wind speed, and X 2 (t) represents the second-order vector of the wind speed, represents the splicing operation on the vector;

[0083] Step 3.2: Input the pan wind speed trajectory obtained in Step 3.1 into a two-layer fully connected network to extract the differential term h x (t) of the wind power controlled by the wind speed, and its simulation process is as follows:

[0084]

[0085] where, W h1 and b h1 are the weight and bias coefficient of the first-layer fully connected layer respectively; W h2 and b h2 are the weight and bias coefficient of the second-layer fully connected layer respectively; relu and tanh represent the rectified linear unit function and the hyperbolic tangent activation function respectively;

[0086] Step 3.3: Combine the wind speed trajectory X(t) and the hidden feature trajectory z(t) to obtain Then input I(t) into the time evolution mechanism module to extract the time inertia information h(t) of the wind power itself; this module is part of the NHDE model, and its specific process is as follows:

[0087]

[0088] where σ represents the sigmoid activation function; p(t) is the update gate, which maps the input to the interval [0,1] through the sigmoid activation function; It is the candidate hidden state, and the change rate of the hidden state over time is simulated by the gated update result after differentiating it from the previous hidden state z(t).

[0089] Step 3.4: Integrate the differential term h x (t) obtained in Step 3.2 and the time inertia information h(t) obtained in Step 3.3 with respect to the wind speed trajectory X(t) and the time term t respectively to simulate the dynamic evolution law of wind power and obtain the hidden feature state z(t) in the target time period:

[0090]

[0091] In the formula, z(t1) represents the hidden feature of the target wind power at time t1, z(t0) is the value at the initial time, t0 and t1 are two adjacent times in the target time period, and the hidden feature states at all adjacent times in the obtained target time period are concatenated in chronological order through the above formula to form z(t).

[0092] In Step 4, the discretization module 2 is used for discretization and concatenation operations. The discretization operation is to discretize the wind speed trajectory X(t) obtained in Step 2 and the continuous trajectory z(t) of the wind power hidden feature obtained in Step 3 in the form of the wind speed tensor {(x0,t0),(x1,t1),...,(x i ,t i ), (x i+1 ,t i+1 ),...,(x n ,t n )}, and the concatenation function is to concatenate the discretized wind speed trajectory and the discretized continuous trajectory z(t) of the wind power hidden feature in the time dimension.

[0093] The fully connected layer module 3 mentioned in Step 4 consists of two fully connected layers.

[0094] The wind speed segmented control module 4 mentioned in Step 5 consists of Figure 1 the multi-layer perceptron model included in, and the usage steps of the multi-layer perceptron model are as follows:

[0095] Step 5.1: Divide the input wind speed into four intervals and set the output power rule, requiring that the wind power time series evolution law learned by the NHDE model is only valid in the wind speed interval of v in < v t < v N , and the wind power output is fixed at 0 when v t < v in or v t > v out , and the wind power output is fixed at 0 when v N > vt < v out is fixed to the rated output power P N ; where v t and P t represent the input wind speed and the output wind power at time t respectively; f(v t ) is a polynomial function with the input wind speed as the independent variable; v in , v N , v out are the cut-in wind speed, the rated wind speed and the cut-out wind speed of the wind turbine respectively;

[0096] Step 5.2: Output the predicted wind power weights {w i,t , i = 1, 2, 3, 4} of the model in the target time period under four wind speed intervals through a multi-layer perceptron with a softmax function in the output layer.

[0097] In step 6, the output layer module 5 consists of two fully connected layers.

[0098] Example:

[0099] The present invention uses a model based on a piecewise-controlled hybrid differential neural network (Neural Piecewise Hybrid Differential Equation, NPHDE) to obtain the temporal characteristics of the short-term power of a wind farm. Specifically: First, the wind speed at discrete time points is continuousized into a wind speed time series trajectory through a cubic spline function, and the hybrid differential neural network is used to parallelly simulate the evolution law of wind power control under the influence of time inertia and continuous wind speed to fully reflect its dynamic evolution mode; then, according to the different control modes of wind power when the wind speed falls into different numerical intervals, a multi-layer perceptron is used to output different predicted wind power weight scores; finally, the short-term predicted value of the wind farm power that simultaneously satisfies the time inertia and the evolution law of wind speed piecewise control is output.

[0100] The system continuousizes the wind speed at discrete time points into a wind speed trajectory by using cubic spline interpolation; parallelly simulates the evolution law of wind power control under the influence of time inertia and continuous wind speed through a hybrid differential neural network model; according to the different control modes of wind power when the wind speed falls into different numerical intervals, a multi-layer perceptron is used to output different predicted wind power weight scores for the short-term power prediction of multiple wind farms.

[0101] In this embodiment, the minute-level wind power generation data of a certain target wind farm for two years is used to perform the day-ahead power prediction of the target wind farm, that is, the prediction result is the wind power value of the wind farm in the next 24 hours. The embodiments of the present invention use the mean absolute percentage error (MAPE) and the root mean square error (RMSE) to evaluate the prediction accuracy, and their calculation formulas are as follows:

[0102]

[0103] Wherein, n is the number of data points of a single sample in the test set; and y t are the predicted value and the actual value of the wind power at time t, respectively; P N is the rated power of the wind farm.

[0104] Figure 2 、 Figure 3 are the wind power prediction curves of each model under the missing rates of 10% and 30% of the input wind speed in the embodiments of the present invention, respectively. It can be seen that whether the missing rate is 10% or 30%, the predicted values of the system of the present invention are closer to the true values at most missing points, indicating that the system has higher prediction accuracy when processing irregular time series data.

[0105] Figure 4 is the error statistics of the ablation experiment. The explanations of each ablation model are as follows:

[0106] 1) NPHDE-D: Fix the wind speed control interval and its weight:

[0107]

[0108] 2) NPHDE-P: Remove the piecewise control strategy of wind speed on wind power;

[0109] 3) NPHDE-O: Replace NHDE with NCDE and ignore the time series inertia of wind power itself;

[0110] 4) NPHDE-C: Remove the control characteristics of wind speed on wind power in the NHDE model.

[0111] From Figure 4 it can be seen that compared with the NPHDE model, the prediction errors of the ablation models increase to varying degrees. Specifically, after adopting the fixed wind speed control interval, the MAPE and RMSE errors of the NPHDE-D model increase by 0.11% and 0.19 MW respectively, indicating that the dynamic piecewise control strategy is more conducive to capturing the dynamic changes at the boundary of the wind speed control interval. After completely removing the piecewise control strategy of wind speed on wind power, the MAPE and RMSE of the NPHDE-P model increase by 0.39% and 0.22 MW respectively, indicating that ignoring the piecewise control strategy of wind speed on power will cause the prediction model to be unable to fully simulate the differential control mode of wind speed on power, resulting in a decrease in prediction accuracy.

[0112] To more intuitively show the effectiveness of the piecewise control strategy, Figure 5The prediction curves of the NPHDE, NPHDE-D, and NPHDE-P models on the test set were compared. From Figure 5 It can be observed that when the wind speed in the time interval from the 1st to the 20th time point is lower than the cut-in wind speed, for the NPHDE (red line) and NPHDE-D models (purple line) with the wind speed segmented control strategy, compared with the NPHDE-P model (blue line), their predicted power is closer to the actual output power of 0. When the wind speed changes in the time interval from the 30th to the 80th time point, resulting in a drastic power fluctuation, both the NPHDE and NPHDE-D models with the wind speed segmented control strategy can quickly adjust the predicted output to match the changing trend of the actual power, but there are obvious differences in the control interval weights of the two. Taking the time interval from the 30th to the 50th time point as an example, when the actual output power of the corresponding wind turbine fluctuates near the rated power (100 MW) due to the dynamic change of the wind speed, the NPHDE-D model with a fixed weight has limited adaptability to the wind speed fluctuation, resulting in most predicted values falling into the rated power and unable to accurately fit the actual power curve. However, the NPHDE model can effectively respond to the wind speed change by dynamically adjusting the wind speed control interval weight, enabling the model to flexibly adjust the power output according to the actual fluctuation of the wind speed.

[0113] From Figure 5 It can also be observed that after removing the control characteristics of the wind speed on the wind power, the MAPE and RMSE errors of the NPHDE-C model increased by 1.03% and 1.11 MW respectively; after ignoring the time series inertia of the wind power itself, the MAPE and RMSE errors of the NPHDE-O model increased by 0.85% and 0.9 MW respectively, indicating that both the wind speed control evolution characteristics and the time series autocorrelation play important roles in revealing the internal dynamic laws of wind power. Combining the experimental conclusions of the segmented control strategy, it can be seen that the wind speed control evolution characteristics contribute the most to the improvement of the model prediction accuracy, followed by the time series autocorrelation, while the contribution degree of the segmented control strategy is relatively low. However, overall, these three mechanisms all play an indispensable role in improving the prediction accuracy of wind power. The combination of the three can more comprehensively simulate the complex evolution law of wind power, thus significantly improving the prediction performance of the model.

[0114] Figure 6 Shows the changes in the MAPE and RMSE errors of the NPHDE model when the number of neurons m increases from 16 to 128. As m increases, RMSE and MAPE generally show the same changing trend, where MAPE fluctuates within 5.63 - 6.90%, with a maximum deviation of 1.27%. When m is small, the information storage in the hidden layer is insufficient, resulting in large prediction errors; as m increases, the capacity of the hidden layer increases, and the model prediction error decreases; when m increases to a certain threshold (100 in this example), the limited training samples may lead to insufficient training of a large number of model parameters, resulting in an increase in errors. The optimal m value selected in this example is 100.

[0115] Figure 7 Statistical chart of the prediction error of the NPHDE model when the model learning rate l varies in the range of [0.0001, 0.01]. Theoretically, using a larger learning rate for the model may skip the optimal solution; while the model with a smaller learning rate may not be able to fully fit the training data. From Figure 8 It can be seen that as the learning rate l decreases from 0.01 to 0.0005, the MAPE error of the model only fluctuates within a small range of 5.63 - 5.67%, the maximum deviation is only 0.04%, and the maximum deviation of RMSE is also only 0.3 MW. This shows that within this range, the change of the learning rate l has limited influence on the model convergence performance and prediction error. However, as the learning rate l further decreases to 0.0001, due to the too small update step, the training process is prone to fall into local optimal solutions or early stop, resulting in a significant increase in the model error. The optimal l value selected for this example is 0.001.

[0116] Figure 8 Shows the change curves of the prediction error of the NPHDE model and the training time of a single sample with respect to Δt, where the abscissa Δt = 1 corresponds to the time step of the original input time series (15 min in this example). Observing the curve, it can be found that as the Euler step size Δt in the hidden state integration increases from 0.1 to 1, the MAPE prediction error of the model shows a trend of first decreasing and then increasing, fluctuating within the range of 5.48 - 6.14%, and the maximum deviation is 0.66%. When Δt = 0.3, the MAPE of the model reaches the lowest value of 5.48%, while the training time of the sample will increase rapidly as the integration step size decreases. When the step size is appropriately reduced, the integration process is more accurate, but too small a step size (such as when Δt = 0.1) will significantly increase the computational amount and computational time of the training process, thereby leading to the accumulation of computational errors and reducing the overall prediction accuracy. Considering both the prediction accuracy and the training time, the optimal integration step size selected for this example is 0.5.

Claims

1. A short-term wind power forecasting system considering wind speed control and power deduction, characterized in that: It comprises a continuation module (1), a discretization module (2), a fully connected layer module (3), a wind speed segmentation control module (4), and an output layer module (5); wherein the output end of the continuation module (1) is connected to the input end of the discretization module (2), the output end of the discretization module (2) is connected to the input end of the fully connected layer module (3), and the wind speed segmentation control module (4) uses the wind speed trajectory obtained in the continuation module (1) to perform segmentation control of the wind speed; the output result of the fully connected layer module (3) and the output result of the wind speed segmentation control module (4) are fused in the time dimension and then input to the input end of the output layer module (5).

2. The system according to claim 1, characterized in that When the system is used, the following steps are taken: Step 1: The historical power data and external wind speed information data of the target wind farm at the minute level are collected through the continuous module (1), and the data are divided into the power vector of the target wind farm and the wind speed tensor at discrete time points; Step 2: The continuation module (1) converts the wind speed tensor in step 1 into a wind speed trajectory X(t) through a cubic spline function, and obtains cubic polynomial wind speed trajectory coefficients; Step 3: The continuation module (1) performs integration calculation and dimension mapping on the cubic polynomial wind speed trajectory coefficients obtained in step 2 to generate a continuous trajectory z(t) containing the hidden characteristics of wind power in the target time period; Step 4: The discretization module (2) concatenates the continuous trajectory z(t) containing the hidden features of wind power in the target time period obtained in step 3 with the wind speed trajectory X(t) obtained in step 2, and inputs the concatenated result into the fully connected layer module (3) to obtain the initial prediction result o(t) of wind power in the target time period; Step 5: Process the wind speed trajectory X(t) obtained in step 2 through the wind speed segmentation control module (4) to obtain a weight vector sequence {w i,t ,i=1,2,3,4}; Step 6: The output layer module (5) converts the initial wind power forecast result o(t) obtained in step 4 into the target time period according to the weight vector sequence {w i,t , i=1,2,3,4}, and input the integrated result to the output layer module (5) to obtain the wind power prediction curve for the target time period The wind power prediction curve can be obtained from the above steps: The wind power prediction curve obtained It is the short-term wind power forecast result.

3. The system according to claim 2, characterized in that In step 1, the wind speed tensor is in the form of {(x0,t0),(x1,t1),...,(x i ,t i ),(x i+1 ,t i+1 ),...,(x n ,t n )}, where (x i ,t i ) represents the i-th time point t i Wind speed value x i .

4. The system according to claim 2, characterized in that In step 2, in each wind speed tensor interval [x i ,x i+1 ], the cubic spline function S i (x) can be expressed as a cubic polynomial: S i (x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 where a i , b i , c i , d i are the coefficients of the cubic polynomial wind speed trajectory.

5. The system according to claim 2, characterized in that The initial prediction result o(t) of wind power in the target time period obtained in step 4 is as follows: Where o(t) is the preliminary prediction result of the target wind power range under the effective wind speed range; represents the first fully connected layer, relu is the linear rectification function, W o1 and b o1 are the weights and bias coefficients of the first fully connected layer, It is the input data of the first fully connected layer; represents the second fully connected layer, tanh is the hyperbolic tangent function, W o2 and b o2 are the weights and bias coefficients of the second fully connected layer, is the output of the first fully connected layer, which serves as the input of the second fully connected layer.

6. The system according to claim 2, characterized in that The specific operation process of step 6 is as follows: Where "·" represents numerical multiplication, (w 1,t 0+w 2,t o t +w 3,t 1+w 4,t 0) represents an integration operation, which represents a fusion operation of the output result of the fully connected layer module (3) and the output result of the wind speed segmented control module (4) in the time dimension; the result of the fusion operation is used as the input of the output layer module (5); represents the wind power value at time t of the final prediction result of the target wind power range under the effective wind speed range; t is the wind power value at time t of the preliminary prediction result of the target wind power range under the effective wind speed range; 0 and 1 are the normalized values ​​at zero power and rated power output, respectively.

7. The system according to claim 1 or 2, characterized in that: The continuation module (1) includes a cubic spline function for continuating discrete wind speeds and an NHDE model for obtaining the hidden characteristic state z(t) of the target time period. When the NHDE model is used to obtain the hidden characteristic state z(t) of the target time period, the following steps are adopted: Step 3.1: Considering that wind power is generally proportional to the cube of wind speed, the first-order differential of wind speed is a quadratic polynomial about wind speed. Therefore, the wind speed power terms not higher than the second order are spliced ​​into the pan-wind speed trajectory. Where X(t) represents the wind speed trajectory, that is, the first direction of the wind speed, X 2 (t) represents the secondary direction quantity of wind speed, Indicates the concatenation operation on the vector; Step 3.2: Substitute the pan wind velocity trajectory obtained in step 3.1 Input a two-layer fully connected network to extract the differential term h of wind power controlled by wind speed x (t), the simulation process is as follows: Where W h1 , b h1 are the weight and bias coefficient of the first fully connected layer; W h2 , b h2 are the weights and bias coefficients of the second fully connected layer respectively; relu and tanh represent the linear rectification function and the hyperbolic tangent activation function respectively; Step 3.3: Combine the wind speed trajectory X(t) and the hidden feature trajectory z(t) to obtain Then I(t) is input into the time evolution mechanism module to extract the time inertia information h(t) of wind power itself; this module is part of the NHDE model, and its specific process is as follows: Where σ represents the sigmoid activation function; p(t) is the update gate, which maps the input to the [0,1] interval through the sigmoid activation function; is the candidate hidden state, and the gated update result after the difference between it and the hidden state z(t) at the previous moment is used to simulate the rate of change of the hidden state over time; Step 3.4: Substitute the differential term h obtained in step 3.2 x (t) and integrate the time inertia information h(t) obtained in step 3.3 with the wind speed trajectory X(t) and the time term t respectively, simulate the dynamic evolution law of wind power, and obtain the hidden characteristic state z(t) in the target time period: In the formula, z(t1) represents the hidden feature of the target wind power at time t1, z(t0) is the initial time value, t0 and t1 are two adjacent times in the target time period. Through the above formula, the hidden feature states of all adjacent times in the target time period are concatenated in chronological order to form z(t).

8. The system according to claim 2, characterized in that In step 4, the discretization module (2) is used to perform discretization and splicing operations, wherein the discretization operation is to convert the wind speed trajectory X(t) obtained in step 2 and the wind power implicit feature continuous trajectory z(t) obtained in step 3 into a wind speed tensor {(x0, t0), (x1, t1), ..., (x i ,t i ),(x i+1 ,t i+1 ),...,(x n ,t n )}, and the splicing function is to splice the discretized wind speed trajectory and the wind power hidden feature continuous trajectory z(t) after discretization in the time dimension.

9. The system according to claim 2, characterized in that The fully connected layer module (3) mentioned in step 4 is composed of two fully connected layers, and in step 6, the output layer module (5) is composed of two fully connected layers.

10. The system according to claim 2, characterized in that The wind speed segmentation control module (4) mentioned in step 5 is composed of the multilayer perceptron model included in FIG1 . The steps of using the multilayer perceptron model are as follows: Step 5.1: Divide the input wind speed into four intervals and set the output power rule, requiring that the wind power time series evolution law learned by the NHDE model only occurs within v in <v t <v N It is valid within the wind speed range of v t <v in or v t >v out In the case of wind power output is fixed to 0, at v N <v t <v out The rated output power P is fixed at N ; where v t and P t Respectively represent the input wind speed and output wind power at time t; f(v t ) is a polynomial function with input wind speed as independent variable; v in 、v N 、v out are the cut-in wind speed, rated wind speed and cut-out wind speed of the wind turbine respectively; Step 5.2: Output the predicted wind power weights {w i,t ,i=1,2,3,4}.

Images