Plateau medium and long term wind power probability prediction method based on adaptive quantile function and CNN-LSTM

Abstract

The invention discloses a plateau medium and long term wind power probability prediction method based on an adaptive quantile function and CNN-LSTM. The method comprises the following steps: 1) using the adaptive quantile function as a wind power probability prediction model; 2) establishing a CNN-LSTM neural network model, extracting spatial-temporal characteristics of the wind power data by using the CNN-LSTM neural network model, and predicting values of to-be-optimized parameters of the adaptive quantile function; 3) optimizing the wind power probability prediction model based on the to-be-optimized parameter value predicted by the CNN-LSTM neural network model; and 4) generating wind power prediction values under different quantiles by using the optimized wind power probability prediction model to obtain a wind power probability prediction result. According to the method, the adaptive quantile function and the CNN-LSTM neural network are combined, and accurate prediction of the medium and long term wind power probability of the plateau is realized.

Description

Technical Field

[0001] The present invention relates to the field of prediction of renewable energy power generation in electric power systems, and in particular to a medium- and long-term wind power probability prediction method in plateaus based on an adaptive quantile function and CNN-LSTM. Background Art

[0002] In recent years, with the rapid development of China's economy, the power load has increased year by year. Due to the shortage of fossil fuels and serious pollution, clean and low-carbon has become the main direction of the new round of global energy transformation. Wind power has become one of the earliest new energy sources developed in various countries in the world due to its advantages such as clean and low-carbon, good economic benefits and short construction period. It is also one of the most valued research directions in the field of new energy power generation. Wind power and photovoltaic power forecasting technology indirectly reduces the impact of power fluctuations on the power grid by estimating the power generation in the future period, optimizes the power flow, and ensures the smooth conduct of power transactions. Grid dispatchers can formulate output plans for other power sources in advance based on the forecast results, and make real-time adjustments based on ultra-short-term forecasts to ensure the stable operation of the power grid. This helps to reduce the demand and cost of backup power and achieve efficient use of wind power and photovoltaic power generation.

[0003] In recent years, the prediction of renewable energy power represented by wind power has become a research hotspot. Traditional wind power predictions such as time series analysis, physical models or statistical models mostly use deterministic methods. These methods can give point predictions of wind power, but cannot effectively deal with the volatility and uncertainty of wind power, resulting in low prediction accuracy in complex environments. At the same time, general machine learning and quantile regression methods are highly dependent on data and cannot fully capture spatiotemporal correlations, and the prediction effect is poor under extreme conditions. However, wind resources in high-altitude areas such as plateaus are highly non-uniform and unstable in space and time, and have characteristics such as low temperature, low pressure, and low air density. Traditional probabilistic prediction methods are difficult to achieve accurate predictions. Therefore, a method is needed to solve the problem of wind power prediction under such special meteorological conditions. Summary of the invention

[0004] The purpose of the present invention is to provide a medium- and long-term plateau wind power probability prediction method based on an adaptive quantile function and CNN-LSTM, comprising the following steps:

[0005] 1) Using adaptive quantile function as wind power probability prediction model;

[0006] 2) Establish a CNN-LSTM neural network model, and use the CNN-LSTM neural network model to extract the spatiotemporal characteristics of wind power data and predict the values ​​of the parameters to be optimized of the adaptive quantile function;

[0007] 3) Optimize the wind power probability prediction model based on the parameter values ​​to be optimized predicted by the CNN-LSTM neural network model;

[0008] 4) Use the optimized wind power probability prediction model to generate wind power prediction values ​​at different quantiles to obtain wind power probability prediction results.

[0009] Furthermore, the CNN-LSTM neural network model includes a CNN network and a LSTM network.

[0010] The CNN network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer;

[0011] The CNN network is used to extract local features of wind power data and input them into the LSTM network;

[0012] The LSTM network predicts the parameters of the adaptive quantile function based on the local characteristics of the wind power data.

[0013] Furthermore, the activation function of the CNN-LSTM neural network model includes a tanh function and a ReLU function, namely:

[0014]

[0015] ReLU(x)=max(0,x) (2)

[0016] Where: x represents the data obtained after the input data is processed by the layer structure, tanh(x) and ReLU(x) represent the output after being processed by the activation function.

[0017] Furthermore, the loss function CRPS(F,y) of the CNN-LSTM neural network model is as follows:

[0018]

[0019] Where: F(y) is the predicted cumulative distribution function, 1 y≥x is the indicator function of the observation, x is the actual observation of the input data, and y is the output.

[0020] Furthermore, in the CNN network,

[0021] The output of the convolutional layer looks like this:

[0022]

[0023] Where: represents the i-th feature map of the l-1th convolutional layer, represents the convolution operation, N i represents the feature atlas of the input layer, represents the weight between the i-th feature map of the l-1th convolutional layer and the j-th feature map of the l-th convolutional layer, Represents the bias between the i-th feature map of the l-1th convolutional layer and the j-th feature map of the l-th convolutional layer; represents the jth feature map of the lth convolutional layer, and f(·) represents the activation function;

[0024] The output of the pooling layer looks like this:

[0025]

[0026] Where: represents the i-th feature map of the l-1th pooling layer, represents the weight coefficient of the jth feature map of the lth pooling layer, down(·) represents the pooling function, Represents the jth feature map of the lth pooling layer;

[0027] The output of the fully connected layer is as follows:

[0028]

[0029] Where: is the output of the jth neuron in the lth layer; is the input data of the fully connected layer.

[0030] Furthermore, the LSTM network sequentially uses a forget gate, an input gate, and an output gate to screen and memorize local features of the wind power data;

[0031] The input of the forget gate includes the output of the previous unit hidden layer and the local feature map of the CNN network output at the current moment.

[0032] The output of the forget gate looks like this:

[0033] f t =σ(W f ·[h t-1 ,x t ]+b f ) (7)

[0034] Where: f t is the final output of the forget gate, W f is the weight, b f is the bias term; h t-1 is the output of the output gate at time t-1; x t is the input;

[0035] The input and output of the input gate are as follows:

[0036] i t =σ(W i ·[ht-1 ,x t ]+b t ) (8)

[0037]

[0038] Where: c t is the output of the input gate, f t is the output of the forget gate, c t-1 is the output of the unit at the previous moment, i t Represents the input of the input gate; W i , W c is the corresponding weight, b t , b c is the bias term, The content obtained after updating the input data features;

[0039] The output of the output gate looks like this:

[0040] o t =σ(W o ·[h t-1 ,x t ]+b o ) (11)

[0041] h t =o t ⊙tanh(c t ) (12)

[0042] In the formula: o t is the intermediate output, h t is the final output of the output gate; W o is the weight, b o is the bias term.

[0043] Furthermore, in step 2), before using the CNN-LSTM neural network model to extract the spatiotemporal features of wind power data, the input data is also normalized;

[0044] The normalized input data is as follows:

[0045]

[0046] Where: x * is the normalized meteorological condition and power generation sample value; x is the meteorological condition and power generation sample value to be normalized; x min is the minimum value of meteorological conditions and power generation samples; x max It is the maximum value of meteorological conditions and power generation sample.

[0047] Furthermore, the input data of the CNN-LSTM neural network model include meteorological conditions and power generation.

[0048] Furthermore, the adaptive quantile function Qu(β) is as follows:

[0049]

[0050] In the formula, is the truncated power function; β∈(0,1) represents the quantile level; D represents the order of the spline; σ k represents the kth node; η0, η1, θ k represents the coefficient of the adaptive quantile function; K represents the number of nodes;

[0051] The constraints of the adaptive quantile function are as follows:

[0052] 0=σ1<σ2<···<σ K <σ K+1 =1 (16)

[0053]

[0054] In the formula, σ K+1 =1;σ1=0;Q u represents the corresponding adaptive quantile function; I k (σ k+1 ) represents a linear constraint on the piecewise defined adaptive quantile function;

[0055] Furthermore, the parameter set Ω to be optimized for the adaptive quantile function is as follows:

[0056] Ω=(η0,η1,χ,δ) (19)

[0057] Among them, the parameter vector to be optimized χ=[χ1,···,χ k-1 ] T ; k =σ k+1 -σ k represents the distance between adjacent nodes; η0 and η1 represent the coefficients of the adaptive quantile function; δ represents the segment distance parameter of the adaptive quantile function, which is used to adjust the smoothness and fitting effect of the quantile function.

[0058] Furthermore, when optimizing the wind power probability prediction model, the optimization objectives are as follows:

[0059]

[0060] Where: y i is the i-th training sample; N is the number of samples; Qu Ωis the adaptive quantile function corresponding to the set Ω. CRPS is the loss function.

[0061] The technical effect of the present invention is unquestionable. The present invention combines the adaptive quantile function with the CNN-LSTM neural network to achieve accurate prediction of the medium- and long-term wind power probability in the plateau. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 The flowchart of the prediction method is shown in Figure 2. DETAILED DESCRIPTION

[0063] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

[0064] Embodiment 1:

[0065] See also Figure 1 , a medium- and long-term plateau wind power probability prediction method based on adaptive quantile function and CNN-LSTM, including the following steps:

[0066] 1) Using adaptive quantile function as wind power probability prediction model;

[0067] 2) Establish a CNN-LSTM neural network model, and use the CNN-LSTM neural network model to extract the spatiotemporal characteristics of wind power data and predict the values ​​of the parameters to be optimized of the adaptive quantile function;

[0068] 3) Optimize the wind power probability prediction model based on the parameter values ​​to be optimized predicted by the CNN-LSTM neural network model;

[0069] 4) Use the optimized wind power probability prediction model to generate wind power prediction values ​​at different quantiles to obtain wind power probability prediction results.

[0070] The CNN-LSTM neural network model includes a CNN network and a LSTM network.

[0071] The CNN network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer;

[0072] The CNN network is used to extract local features of wind power data and input them into the LSTM network;

[0073] The LSTM network predicts the parameters of the adaptive quantile function based on the local characteristics of the wind power data.

[0074] The activation functions of the CNN-LSTM neural network model include the tanh function and the ReLU function, namely:

[0075]

[0076] ReLU(x)=max(0,x) (2)

[0077] Where: x represents the data obtained after the input data is processed by the layer structure, tanh(x) and ReLU(x) represent the output after being processed by the activation function.

[0078] The loss function CRPS(F,y) of the CNN-LSTM neural network model is as follows:

[0079]

[0080] Where: F(y) is the predicted cumulative distribution function, 1 y≥x is the indicator function of the observation, x is the actual observation of the input data, and y is the output.

[0081] In the CNN network, the output of the convolutional layer is as follows:

[0082]

[0083] Where: represents the i-th feature map of the l-1th convolutional layer, represents the convolution operation, N i represents the feature atlas of the input layer, represents the weight between the i-th feature map of the l-1th convolutional layer and the j-th feature map of the l-th convolutional layer, Represents the bias between the i-th feature map of the l-1th convolutional layer and the j-th feature map of the l-th convolutional layer; represents the jth feature map of the lth convolutional layer, and f(·) represents the activation function;

[0084] The output of the pooling layer looks like this:

[0085]

[0086] Where: represents the i-th feature map of the l-1th pooling layer, represents the weight coefficient of the jth feature map of the lth pooling layer, down(·) represents the pooling function, Represents the jth feature map of the lth pooling layer;

[0087] The output of the fully connected layer is as follows:

[0088]

[0089] Where: is the output of the jth neuron in the lth layer; is the input data of the fully connected layer.

[0090] The LSTM network sequentially uses a forget gate, an input gate, and an output gate to screen and memorize local features of wind power data;

[0091] The input of the forget gate includes the output of the previous unit hidden layer and the local feature map of the CNN network output at the current moment.

[0092] The output of the forget gate looks like this:

[0093] f t =σ(W f ·[h t-1 ,x t ]+b f ) (7)

[0094] Where: f t is the final output of the forget gate, W f is the weight, b f is the bias term; h t-1 is the output of the output gate at time t-1; x t is the input;

[0095] The input and output of the input gate are as follows:

[0096] i t =σ(W i ·[h t-1 ,x t ]+b t ) (8)

[0097]

[0098] Where: c t is the output of the input gate, f t is the output of the forget gate, c t-1 is the output of the unit at the previous moment, i t Represents the input of the input gate; W i , W c is the corresponding weight, b t , b c is the bias term, The content obtained after updating the input data features;

[0099] The output of the output gate looks like this:

[0100] o t =σ(W o ·[h t-1,x t ]+b o ) (11)

[0101] h t =o t ⊙tanh(c t ) (12)

[0102] In the formula: o t is the intermediate output, h t is the final output of the output gate; W o is the weight, b o is the bias term.

[0103] In step 2), before using the CNN-LSTM neural network model to extract the spatiotemporal characteristics of wind power data, the input data is also normalized;

[0104] The normalized input data is as follows:

[0105]

[0106] Where: x * is the normalized meteorological condition and power generation sample value; x is the meteorological condition and power generation sample value to be normalized; x min is the minimum value of meteorological conditions and power generation samples; x max It is the maximum value of meteorological conditions and power generation sample.

[0107] The input data of the CNN-LSTM neural network model include meteorological conditions and power generation.

[0108] The adaptive quantile function Qu(β) is as follows:

[0109]

[0110] In the formula, is the truncated power function; β∈(0,1) represents the quantile level; D represents the order of the spline; σ k represents the kth node; η0, η1, θ k represents the coefficient of the adaptive quantile function; K represents the number of nodes;

[0111] The constraints of the adaptive quantile function are as follows:

[0112] 0=σ1<σ2<···<σ K <σ K+1 =1 (16)

[0113]

[0114] In the formula, σK+1 =1;σ1=0;Q u represents the corresponding adaptive quantile function; I k (σ k+1 ) represents a linear constraint on the piecewise defined adaptive quantile function;

[0115] The parameter set Ω to be optimized for the adaptive quantile function is as follows:

[0116] Ω=(η0,η1,χ,δ) (19)

[0117] Among them, the parameter vector to be optimized χ=[χ1,···,χ k-1 ] T ; k =σ k+1 -σ k represents the distance between adjacent nodes; η0 and η1 represent the coefficients of the adaptive quantile function; δ represents the segment distance parameter of the adaptive quantile function, which is used to adjust the smoothness and fitting effect of the quantile function.

[0118] When optimizing the wind power probability prediction model, the optimization objectives are as follows:

[0119]

[0120] Where: y i is the i-th training sample; N is the number of samples; Qu Ω is the adaptive quantile function corresponding to the set Ω. CRPS is the loss function.

[0121] Embodiment 2:

[0122] A medium- and long-term plateau wind power probability prediction method based on adaptive quantile function and CNN-LSTM includes the following steps:

[0123] 1) Using adaptive quantile function as wind power probability prediction model;

[0124] 2) Establish a CNN-LSTM neural network model, and use the CNN-LSTM neural network model to extract the spatiotemporal characteristics of wind power data and predict the values ​​of the parameters to be optimized of the adaptive quantile function;

[0125] 3) Optimize the wind power probability prediction model based on the parameter values ​​to be optimized predicted by the CNN-LSTM neural network model;

[0126] 4) Use the optimized wind power probability prediction model to generate wind power prediction values ​​at different quantiles to obtain wind power probability prediction results.

[0127] Embodiment 3:

[0128] A method for predicting medium- and long-term wind power probability in plateau based on an adaptive quantile function and CNN-LSTM, the technical content of which is the same as that of Example 2, and further, the CNN-LSTM neural network model includes a CNN network and a LSTM network.

[0129] The CNN network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer;

[0130] The CNN network is used to extract local features of wind power data and input them into the LSTM network;

[0131] The LSTM network predicts the parameters of the adaptive quantile function based on the local characteristics of the wind power data.

[0132] Embodiment 4:

[0133] A method for predicting medium- and long-term wind power probability in plateau based on an adaptive quantile function and CNN-LSTM, the technical content of which is the same as any one of Embodiments 2-3, and further, the activation function of the CNN-LSTM neural network model includes a tanh function and a ReLU function, namely:

[0134]

[0135] ReLU(x)=max(0,x) (2)

[0136] Where: x represents the data obtained after the input data is processed by the layer structure, tanh(x) and ReLU(x) represent the output after being processed by the activation function.

[0137] Embodiment 5:

[0138] A method for predicting medium- and long-term wind power probability in plateau based on an adaptive quantile function and CNN-LSTM, the technical content of which is the same as any one of Embodiments 2-4, and further, the loss function of the CNN-LSTM neural network model is as follows:

[0139]

[0140] Where: F(y) is the predicted cumulative distribution function, 1 y≥x is the indicator function of the observation, and x is the actual observation value of the input data.

[0141] Embodiment 6:

[0142] A method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM, the technical content of which is the same as any one of Embodiments 2-5, and further, in the CNN network, the output of the convolutional layer is as follows:

[0143]

[0144] Where: represents the i-th feature map of the l-1th convolutional layer, represents the convolution operation, N i represents the feature atlas of the input layer, represents the weight between the i-th feature map of the l-1th convolutional layer and the j-th feature map of the l-th convolutional layer, Represents the bias between the i-th feature map of the l-1th convolutional layer and the j-th feature map of the l-th convolutional layer; represents the jth feature map of the lth convolutional layer, and f(·) represents the activation function;

[0145] The output of the pooling layer looks like this:

[0146]

[0147] Where: represents the i-th feature map of the l-1th pooling layer, represents the weight coefficient of the jth feature map of the lth pooling layer, down(·) represents the pooling function, Represents the jth feature map of the lth pooling layer;

[0148] The output of the fully connected layer is as follows:

[0149]

[0150] Where: is the output of the jth neuron in the lth layer; is the input data of the fully connected layer.

[0151] Embodiment 7:

[0152] A method for predicting medium- and long-term wind power probability in a plateau based on an adaptive quantile function and CNN-LSTM, the technical content of which is the same as any one of Embodiments 2-6, and further, the LSTM network sequentially uses a forget gate, an input gate, and an output gate to screen and memorize local features of wind power data;

[0153] The input of the forget gate includes the output of the previous unit hidden layer and the local feature map of the CNN network output at the current moment.

[0154] The output of the forget gate looks like this:

[0155] f t =σ(W f ·[h t-1 ,x t ]+b f ) (7)

[0156] Where: ft is the final output of the forget gate, W is the weight, and b is the bias term;

[0157] The input and output of the input gate are as follows:

[0158] i t =σ(W i ·[h t-1 ,x t ]+b t ) (8)

[0159]

[0160] Where: c t is the output of the input gate, f t is the output of the forget gate, c t-1 is the output of the unit at the previous moment, i t Represents the input of the input gate; W i , W c is the corresponding weight, b t , b c is the bias term, The content obtained after updating the input data features;

[0161] The output of the output gate looks like this:

[0162] o t =σ(W o ·[h t-1 ,x t ]+b o ) (11)

[0163] h t =o t ⊙tanh(c t ) (12)

[0164] In the formula: o t is the intermediate output, h t is the final output of the output gate; W o is the weight, b o is the bias term.

[0165] Embodiment 8:

[0166] A medium- and long-term plateau wind power probability prediction method based on an adaptive quantile function and CNN-LSTM, the technical content of which is the same as any one of Embodiments 2-7, and further, in step 2), before using the CNN-LSTM neural network model to extract the spatiotemporal characteristics of wind power data, the input data is also normalized;

[0167] The normalized input data is as follows:

[0168]

[0169] Where: x * is the normalized meteorological condition and power generation sample value; x is the meteorological condition and power generation sample value to be normalized; x min is the minimum value of meteorological conditions and power generation samples; x max It is the maximum value of meteorological conditions and power generation sample.

[0170] Embodiment 9:

[0171] A method for medium- and long-term wind power probability prediction in plateau based on an adaptive quantile function and CNN-LSTM, the technical content of which is the same as any one of Examples 2-8, and further, the input data of the CNN-LSTM neural network model includes meteorological conditions and power generation power.

[0172] Embodiment 10:

[0173] A method for predicting medium- and long-term wind power probability in plateau based on an adaptive quantile function and CNN-LSTM, the technical content of which is the same as any one of Embodiments 2-9, and further, the adaptive quantile function is as follows:

[0174]

[0175] In the formula, is the truncated power function; β∈(0,1) represents the quantile level; D represents the order of the spline; σ k represents the kth node; η0, η1, θ k represents the coefficient of the adaptive quantile function; K represents the number of nodes;

[0176] The constraints of the adaptive quantile function are as follows:

[0177] 0=σ1<σ2<···<σ K <σ K+1 =1 (16)

[0178]

[0179] In the formula, σ K+1 =1;σ1=0;Q u represents the corresponding adaptive quantile function; I k (σ k+1 ) represents a linear constraint on the piecewise defined adaptive quantile function;

[0180] Embodiment 11:

[0181] A method for predicting medium- and long-term wind power probability in plateau based on an adaptive quantile function and CNN-LSTM, the technical content of which is the same as any one of Embodiments 2-10, and further, a set of parameters to be optimized for the adaptive quantile function is as follows:

[0182] Ω=(η0,η1,χ,δ) (19)

[0183] Among them, the parameter vector to be optimized χ=[χ1,···,χ k-1 ] T ; k =σ k+1 -σ k represents the distance between adjacent nodes; η0 and η1 represent the coefficients of the adaptive quantile function; δ represents the segment distance parameter of the adaptive quantile function, which is used to adjust the smoothness and fitting effect of the quantile function.

[0184] Embodiment 12:

[0185] A medium- and long-term plateau wind power probability prediction method based on an adaptive quantile function and CNN-LSTM, the technical content of which is the same as any one of Embodiments 2-11. Further, when optimizing the wind power probability prediction model, the optimization target is as follows:

[0186]

[0187] Where: y i is the ith training sample.

[0188] Embodiment 13:

[0189] A medium- and long-term wind power probability prediction method for plateau based on adaptive quantile function and CNN-LSTM, the steps are as follows:

[0190] 1) Establish a CNN-LSTM neural network model to extract the spatiotemporal characteristics of wind power data and obtain the parameters of the relevant adaptive quantile function;

[0191] 1.1) Perform data preprocessing. Before using neural networks for prediction, it is necessary to eliminate the error caused by the dimension, and at the same time reduce the impact of the dimension. Convert the input data into scalar values, use maximum value normalization for processing, and map the data to [0,1], as shown in the following formula.

[0192]

[0193] Where: x * is the normalized sample value of meteorological conditions, power generation, etc.; x is the sample value of meteorological conditions, power generation, etc. to be normalized; x min is the minimum value of each sample of meteorological conditions, power generation, etc.; x maxThese are the maximum values ​​of various samples such as meteorological conditions and power generation.

[0194] 1.2) Select tanh function and ReLU function as activation functions. The calculation formulas of the two functions are shown below.

[0195]

[0196] ReLU(x)=max(0,x) (3)

[0197] Where: x represents the data obtained after the actual input such as meteorological conditions and power generation is processed by the layer structure, and tanh(x) and ReLU(x) represent the output after being processed by the activation function.

[0198] The continuous ranking probability score (CRPS) is selected as the loss function. A weight is assigned to each prediction source and the CRPS value of each prediction source is calculated based on the assigned weight. The larger the weight, the greater the impact on the final result. Compared with other loss functions such as quantile loss, CRPS can measure the compatibility of the predicted distribution and the actual data, which is more comprehensive. The calculation formula is as follows.

[0199]

[0200] Where: F(y) is the predicted cumulative distribution function, 1 y≥x is the indicator function of the observed value, and x is the actual observed value such as meteorological conditions and power generation.

[0201] 1.3) Use convolutional neural network (CNN) to extract local patterns and spatial dependencies in input data. The local features of data such as wind speed and temperature can be effectively captured through convolution operations, and the extracted features will be used as input for subsequent LSTM networks. The convolutional neural network consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Convolution operations are performed in the convolutional layer to extract the deep features of the data. The calculation formula is shown below.

[0202]

[0203] Where: I l1i The i-th feature graph represents the meteorological conditions, power generation and other data of the l-1th layer, represents the convolution operation, N i The feature atlas represents the weather conditions, power generation and other data of the input layer, w li,j The weight between the i-th feature graph of the meteorological conditions, power generation and other data of the l-1th layer and the j-th feature graph of the meteorological conditions, power generation and other data of the lth layer, The offset between the i-th feature graph of the weather conditions, power generation and other data of the l-1th layer and the j-th feature graph of the weather conditions, power generation and other data of the lth layer, represents the j-th feature map of the weather conditions, power generation and other data of the l-th layer, and f(·) represents the activation function used.

[0204] Afterwards, the pooling layer is used for pooling calculation, and the feature data is sampled for dimensionality reduction, which reduces the number of network parameters, reduces the amount of calculation, improves the calculation efficiency, and can realize the secondary extraction of data features such as wind speed and temperature, which can reduce the overfitting phenomenon. The calculation formula of the pooling layer is as follows.

[0205]

[0206] Where: The i-th feature graph represents the meteorological conditions, power generation and other data of the l-1th layer, represents the weight coefficient of the jth feature map of the meteorological conditions, power generation and other data of the lth layer, down(·) represents the pooling function, The jth feature graph represents the weather conditions, power generation and other data of the lth layer.

[0207] After that, after being processed by the convolutional layer and the pooling layer, all the input feature information is summarized using the fully connected layer. Each neuron in the layer is fully connected to the neurons in the upper layer to facilitate subsequent feature classification, prediction and other operations. The calculation formula of this layer is shown below.

[0208]

[0209] Where: is the jth neuron in the lth layer.

[0210] 1.4) Use long short-term memory network to capture long-term dependencies in time series. The input of LSTM is the feature map output by CN. LSTM further. In LSTM, multiple steps of calculation and transmission are required. The three gate structures of forget gate, input gate and output gate are used to filter and memorize information, thereby realizing long-term memory capability. The input data is first filtered by the forget gate to select the information to be retained and the information to be forgotten. The calculation formula is shown below.

[0211] f t =σ(W f ·[h t-1 ,x t ]+b f ) (8)

[0212] Where: f tis the final output of the forget gate, W is the weight, and b is the bias term. The input of the forget gate is the output of the previous unit hidden layer and the feature map of the meteorological data and power data extracted by the CNN at the current moment. The calculation result determines the C t-1 The retention ratio, that is, how much information needs to be retained from the output of the neuron at the previous moment.

[0213] After that, the cell state is updated through the input gate, and the activation function is used to update the information to obtain two parts of data, which are combined to complete the update of the entire cell state. The calculation process is shown below.

[0214] i t =σ(W i ·[h t-1 ,x t ]+b t ) (9)

[0215]

[0216] Where: i t is the calculation structure of the input gate, W is the corresponding weight, b is the bias term, This is the content obtained after updating the previous meteorological data and other features. After that, the network structure is updated by combining the output of the previous moment and the content of the forget gate. The calculation process is as follows.

[0217]

[0218] Where: C t is the calculation result, f t is the output of the forget gate, C t-1 is the output of the unit at the previous moment, i t Represents the input of the input gate.

[0219] After that, the output gate finally determines the final state of the output. The output of the output gate is obtained by multiplying the matrix formed by connecting the control unit state at the previous moment and the input at the current moment with the weight. t The output content is calculated by tanh to obtain the final output. The calculation process is as follows.

[0220] o t =σ(W o ·[h t-1 ,x t ]+b o ) (12)

[0221] h t =o t ⊙tanh(c t ) (13)

[0222] Where: Ot is the output gate output, h t is the final output, which is the result of comprehensive processing of the input wind speed, characteristics, etc. The result is the parameter required for the subsequent quantile function calculation, which is used to further generate prediction values ​​at different probability levels. W is the weight and b is the bias term.

[0223] 2) Use the adaptive quantile function to perform quantile prediction, generate wind power predictions under different quantiles, and obtain the wind power prediction interval.

[0224] 2.1) Introduce the adaptive quantile function and use the quantile function (Q F ) is used as a prediction model to determine the distribution of wind power generation, and the quantile function of the variable Y is used to represent the future wind power generation. The specific expression is shown below.

[0225]

[0226] Where: β∈(0,1) represents the quantile level (QL), and represents the cumulative distribution function (CDF). The probability density function is defined as F Y (z), and the inverse function of the cumulative distribution function is represented by Q:[0,1]→R, Q L The mapping is represented by β. In addition, y (βi) It is given by the formula The specific βi-quantile value calculated. At the same time, β i ∈(0,1) obtains a specific value through the prediction set. The quantile function is a function of the β quantile level and can be used for wind power prediction.

[0227] 2.2) Use spline function to approximate the quantile function Q of wind power F , we get the spline quantile function (SQ F ). The D-order spline quantile function describing K knots is shown below.

[0228]

[0229] Where: K represents the number of nodes, D represents the order of the spline. d and θ k It's SQ F The coefficient of .

[0230] At the same time, the truncated power function The definition of is as follows.

[0231]

[0232] Where: krepresents the kth node. This is the simplest linear SQ F form.

[0233] Thanks to Q F There is a lack of flatness at each node, so a quadratic smoothing spline function (QQ F ), SQ F The D of is set to 2, so that the continuity is maintained at each node and smoothness is also shown. The specific expression is as follows.

[0234]

[0235] But at the same time, for any parameter set, QQ F may not be valid, so additional constraints are needed. The first constraint is shown below.

[0236] 0=σ1<σ2<···<σ K <σ K+1 =1 (18)

[0237] The first constraint can make all nodes sorted correctly. That is, the upper and lower boundary nodes are set to be σ K+1 = 1 and σ1 = 0, the domain [0,1] is split into K+1 nodes, creating K segment intervals. The second constraint is as follows.

[0238]

[0239] The second constraint requires QQ F The first-order derivative of is greater than or equal to 0, so that the derived quantile function increases rather than decreases. Combining the two constraints, QQ F The expression is as follows.

[0240] χ k =σ k+1 -σ k (20)

[0241] The distance of each node is denoted by χ. Considering the first restriction, χ k The new variable must satisfy χ k ≥0 and Therefore, the modified QQ F The parameters become χ=[χ1,···,χ k-1 ]T. Therefore, σ k Can be used to express Considering the second constraint, β∈[σ k ,σ k+1 The calculation formula for the kth part of ] is as follows.

[0242]

[0243] Where: I k (σ k )=I k-1 (σ k ). Therefore the third constraint is as follows.

[0244]

[0245] Once k = 0 and I0(σ1) = γ1, the new parameter τ can be used for other k ≥ 1 k To replace I k (σ k+1 ). Therefore, we can use τ k and γ1 determine QQ F The initial parameter θ k τ k and γ1 are used to define the constraints of the initial and subsequent nodes. The results after iteration are as follows:

[0246]

[0247] Finally, QQ F The parameter set requiring 2K+1 parameters is shown below.

[0248] Ω=(η0,η1,χ,δ), (25)

[0249] In summary, γ0∈R,γ1∈R + ,χ∈[0,1] K-1 ,τ∈R K+ , K is the number of nodes, is QQ F The hyperparameters of γ0∈R,γ1∈R are used to describe the parameter set Ω and substitute it into the calculation formula to realize the probability prediction of wind power. + It is used to define the constraints of the initial and subsequent nodes. The results after iteration are shown in formulas (23)-(24).

[0250] 2.3) Optimize the parameters obtained by CNN-LSTM and use the optimized parameters to calculate the final probability prediction interval. Parameter evaluation is achieved by minimizing the average CRPS of all training samples to optimize the model. The calculation method is as follows.

[0251]

[0252] Where: The i-th training sample is defined as y i .

[0253] 2.4) Use the CNN-LSTM neural network to continuously update the optimization parameters to achieve adaptive update of the quantile function parameters. Substitute the obtained optimization parameters into formula (17) in 2.2) to calculate and obtain the corresponding wind power probability prediction results.

Claims

1. A medium- and long-term plateau wind power probability prediction method based on adaptive quantile function and CNN-LSTM, characterized in that: The following steps are involved: 1) Using adaptive quantile function as wind power probability prediction model; 2) Establish a CNN-LSTM neural network model, and use the CNN-LSTM neural network model to extract the spatiotemporal characteristics of wind power data and predict the values ​​of the parameters to be optimized of the adaptive quantile function; 3) Optimize the wind power probability prediction model based on the parameter values ​​to be optimized predicted by the CNN-LSTM neural network model; 4) Use the optimized wind power probability prediction model to generate wind power prediction values ​​at different quantiles to obtain wind power probability prediction results.

2. The method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM neural network according to claim 1 is characterized in that: The CNN-LSTM neural network model includes a CNN network and a LSTM network. The CNN network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; The CNN network is used to extract local features of wind power data and input them into the LSTM network; The LSTM network predicts the parameters of the adaptive quantile function based on the local characteristics of the wind power data.

3. The method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM neural network according to claim 2 is characterized in that: The activation functions of the CNN-LSTM neural network model include the tanh function and the ReLU function, namely: ReLU(x)=max(0,x) (2) Where: x represents the data obtained after the input data is processed by the layer structure, tanh(x) and ReLU(x) represent the output after being processed by the activation function.

4. The method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM neural network according to claim 2 is characterized in that: The loss function CRPS(F,y) of the CNN-LSTM neural network model is as follows: Where: F(y) is the predicted cumulative distribution function, 1 y≥x It is the indicator function of the observation value, x is the actual observation value of the input data; y is the output.

5. The method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM neural network according to claim 2 is characterized in that: In the CNN network, the output of the convolutional layer is as follows: Where: represents the i-th feature map of the l-1th convolutional layer, represents the convolution operation, N i represents the feature atlas of the input layer, represents the weight between the i-th feature map of the l-1th convolutional layer and the j-th feature map of the l-th convolutional layer, Represents the bias between the i-th feature map of the l-1th convolutional layer and the j-th feature map of the l-th convolutional layer; represents the jth feature map of the lth convolutional layer, and f(·) represents the activation function; The output of the pooling layer looks like this: Where: represents the i-th feature map of the l-1th pooling layer, represents the weight coefficient of the jth feature map of the lth pooling layer, down(·) represents the pooling function, Represents the jth feature map of the lth pooling layer; The output of the fully connected layer is as follows: Where: is the output of the jth neuron in the lth layer; is the input data of the fully connected layer.

6. The method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM neural network according to claim 2 is characterized in that: The LSTM network sequentially uses a forget gate, an input gate, and an output gate to screen and memorize local features of wind power data; Among them, the input of the forget gate includes the output of the hidden layer of the previous unit and the local feature map output by the CNN network at the current moment; The output of the forget gate looks like this: f t =σ(W f ·[h t-1 ,x t ]+b f ) (7) Where: f t is the final output of the forget gate, W f is the weight, b f is the bias term; h t-1 is the output of the output gate at time t-1; x t For input data; The input and output of the input gate are as follows: Where: c t is the output of the input gate, f t is the output of the forget gate, c t-1 is the output of the unit at the previous moment, i t Represents the input of the input gate; W i , W c is the corresponding weight, b t 、b c is the bias term, The content obtained after updating the input data features; The output of the output gate looks like this: the t =σ(W o ·[h t-1 ,x t ]+b o ) (11) h t =o t ⊙tanh(c t ) (12) In the formula: o t is the intermediate output, h t is the final output of the output gate; W o is the weight, b o is the bias term.

7. The method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM neural network according to claim 1 is characterized in that: In step 2), before using the CNN-LSTM neural network model to extract the spatiotemporal characteristics of wind power data, the input data is also normalized; The normalized input data is as follows: Where: x * is the normalized meteorological condition and power generation sample value; x is the meteorological condition and power generation sample value to be normalized; x min is the minimum value of meteorological conditions and power generation samples; x max It is the maximum value of meteorological conditions and power generation sample.

8. The method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM neural network according to claim 1 is characterized in that: The adaptive quantile function Qu(β) is as follows: In the formula, is the truncated power function; β∈(0,1) represents the quantile level; D represents the order of the spline; σ k represents the kth node; η0, η1, θ k represents the coefficient of the adaptive quantile function; K represents the number of nodes; The constraints of the adaptive quantile function are as follows: 0=σ1<σ2<···<σ K <s K+1 =1 (16) In the formula, σ K+1 =1;σ1=0;Q u represents the corresponding adaptive quantile function; I k (σ k+1 ) represents a linear constraint on the piecewise defined adaptive quantile function.

9. The method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM neural network according to claim 1 is characterized in that: The parameter set Ω to be optimized for the adaptive quantile function is as follows: Ω=(η0,η1,χ,δ) (19) Among them, the parameter vector to be optimized χ=[χ1,···,χ k-1 ] T ; k =σ k+1 -σ k represents the distance between adjacent nodes; η0 and η1 represent the coefficients of the adaptive quantile function; δ represents the segment distance parameter of the adaptive quantile function.

10. The method for predicting medium- and long-term wind power probability in plateau based on adaptive quantile function and CNN-LSTM neural network according to claim 1, characterized in that: When optimizing the wind power probability prediction model, the optimization objectives are as follows: Where: y i is the i-th training sample; N is the number of samples; Qu Ω is the adaptive quantile function corresponding to the set Ω. CRPS is the loss function.

Images