Photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM
By introducing dynamic feature extraction and CNN-BiLSTM model in photovoltaic power prediction, the problem of low photovoltaic power prediction accuracy in the prior art is solved, and more efficient and accurate photovoltaic power prediction is achieved.
Patent Information
- Application Number
- CN202411603489.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-11
- Publication Date
- 2025-03-07
AI Technical Summary
When existing photovoltaic power prediction methods process strong volatility and nonlinear photovoltaic historical data, the prediction accuracy is not high, and adding auxiliary information such as environmental factors will increase the prediction cost.
The photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM is adopted to extract the photovoltaic power generation trend characteristics through moving average and Bollinger Band indexes, and incorporate them into the neural network model. Combining the advantages of CNN and BiLSTM, nonlinear local features and bidirectional features of time sequence data between multi-dimensional input data are extracted.
The accuracy of photovoltaic power prediction is improved, the uncertainty of prediction results is reduced, and the prediction cost is reduced through dynamic feature extraction indicators.
Smart Images
Figure CN119578917A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of photovoltaic power prediction, and in particular relates to a photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM. Background Art
[0002] With the rapid expansion of photovoltaic power generation, how to effectively predict its output power has become the focus of attention in the industry.
[0003] Photovoltaic power generation, as a new energy power generation technology, is mainly divided into two forms: centralized and distributed. Among them, distributed photovoltaic has the advantages of high energy utilization, low environmental pollution, flexible local consumption, and low abandonment rate. Compared with traditional power generation methods, photovoltaic technology has many advantages, but it still has some problems. For example, due to its very complex historical data, numerous data points, and wide coverage, the cost of photovoltaic data statistics is very high and the efficiency is not ideal, which limits the development of photovoltaic technology. At the same time, photovoltaic power generation has strong randomness and volatility, and its output power is affected by many factors, including solar radiation, ambient temperature, cloud cover, etc., which makes photovoltaic prediction a challenging task. The strong intermittent and random nature of photovoltaic power generation has a great impact on power system planning and power supply quality, and this impact will become more prominent as the photovoltaic grid-connected capacity increases. Therefore, in the face of a high proportion of distributed photovoltaic access to the grid, accurately predicting the short-term power generation of distributed photovoltaics is of great significance to the operation and dispatch of the grid.
[0004] At present, there are three main methods for short-term prediction of photovoltaic power: traditional mathematical statistics, machine learning, and deep learning. The commonly used traditional mathematical statistics method model principle is clear, simple, and efficient, and can reflect the time series of input data, but it is not suitable for the prediction of photovoltaic historical data with strong volatility and nonlinearity. The commonly used machine learning algorithm model can take into account the nonlinear relationship between various influencing factors and photovoltaic power generation, but ignores the time series dependence of photovoltaic historical data itself, resulting in a decrease in prediction accuracy and destroying the integrity of the time series data.
[0005] With the development of deep learning, the bidirectional long short-term memory neural network (BiLSTM) came into being. It consists of two layers of forward and reverse networks, which can simultaneously consider the contextual information in both the forward and backward directions, effectively improving the modeling ability of sequence data, thereby improving the accuracy of model prediction. The existing method usually combines multiple single models to form a combined prediction model, and better predicts photovoltaic power generation by combining the advantages of different models. In order to further improve the prediction performance of the model, many algorithms also choose to improve the input data of the model by decomposing historical data. However, due to the strong volatility and instability of photovoltaic power generation data, even after decomposition, it may still lead to poor prediction results, and adding auxiliary information such as environmental factors increases the cost of prediction. Summary of the invention
[0006] To this end, in view of the defects and shortcomings of the above existing technologies, the present invention proposes a photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM, which introduces moving average and Bollinger Band indicators to characterize and extract photovoltaic power generation trend characteristics, and integrates the effective potential information contained in them into the algorithm model input data. The advantages of CNN and BiLSTM neural networks are integrated to extract nonlinear local features between multidimensional input data, and bidirectional feature extraction of time series data is performed to output photovoltaic power generation prediction results.
[0007] The present invention provides theoretical and algorithmic tool support for the research on short-term photovoltaic power prediction of distributed photovoltaic power stations by constructing a photovoltaic power prediction model based on dynamic feature extraction and CNN-BiLSTM.
[0008] The technical solution specifically adopted by the present invention to solve the technical problem is:
[0009] A photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM: The photovoltaic power generation trend characteristics are extracted by moving average and Bollinger Band indicators to obtain dynamic feature parameters, and the nonlinear local features between multi-dimensional input data are extracted by fusing CNN and BiLSTM neural networks. The bidirectional feature extraction of time series data is performed to output the photovoltaic power generation prediction result.
[0010] Furthermore, the dataset is constructed as follows:
[0011] Firstly, the historical data of photovoltaic power and influencing factors are preprocessed, including: removing data collection points during abnormal periods and using the Pearson coefficient to screen strongly correlated influencing factors, to obtain an effective data set after dimensionality reduction;
[0012] Then the photovoltaic power history data and effective influencing factor data are normalized and divided into training set and test set;
[0013] Calculate the dynamic characteristic parameters of the moving average and Bollinger Bands indicators of the effective historical data series of photovoltaic power;
[0014] According to the calculated historical data moving average and Bollinger Bands indicator dynamic feature extraction parameters, they are normalized and added to the data set; data sequences of all dimensions are input into the fusion CNN and BiLSTM neural network for training and prediction.
[0015] Furthermore, according to a photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM according to claim 1, it is characterized by using Bollinger Bands and moving averages to analyze the form of fluctuation data to reflect the dynamic characteristics of photovoltaic power generation corresponding to different forms.
[0016] Furthermore, the calculation formula of the moving average indicator is:
[0017]
[0018] Where: p k Represents the photovoltaic power output fluctuation value data for the n days before the i-th day.
[0019] Furthermore, the moving average is used to construct the middle line of the Bollinger Bands indicator to eliminate the interference factors in the original sequence. The calculation is as follows:
[0020]
[0021] Where: Indicates the standard deviation of the photovoltaic power output data for the previous n days on the i-th day; BOLL UP , BOLL DN They represent the upper and lower lines of the Bollinger Bands indicator respectively, and α represents the adjustment parameter of the bandwidth range between the upper and lower lines of the Bollinger Bands indicator.
[0022] Furthermore, the fusion CNN and BiLSTM neural network is composed of a convolutional neural network CNN and a bidirectional long short-term memory network BiLSTM; the predicted value is made to track the actual photovoltaic power generation through the CNN model, and the long-term dependency in the sequence is learned through the BiLSTM network.
[0023] And, a photovoltaic power prediction system based on dynamic feature extraction and CNN-BiLSTM, comprising:
[0024] Dynamic feature parameter extraction module, used to extract photovoltaic power generation trend characteristics through moving average and Bollinger Band indicators to obtain dynamic feature parameters;
[0025] The prediction module is used to extract nonlinear local features between multi-dimensional input data by fusing CNN and BiLSTM neural networks, and to output photovoltaic power generation prediction results after performing bidirectional feature extraction on time series data.
[0026] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of a photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM are implemented as described above.
[0027] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM as described above.
[0028] Compared with the prior art, the present invention and its preferred solution introduce the CNN model to make the predicted value better track the actual photovoltaic power generation, and at the same time, it can reduce the uncertainty of the prediction result through its powerful feature extraction ability. The BiLSTM network can learn the long-term dependencies in the sequence and make up for the shortcomings of the convolutional neural network CNN in this regard. It is difficult to extract the temporal correlation and change law of features simultaneously when CNN and BiLSTM are used separately. The SMA and BOLL dynamic feature extraction indicators are introduced to provide more effective influencing factor data input for the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0030] Figure 1 This is a BOLL-CNN-BiLSTM photovoltaic prediction flow chart of an embodiment of the present invention;
[0031] Figure 2 This is the structure diagram of the LSTM neural network;
[0032] Figure 3 It is the BiLSTM neural network structure. DETAILED DESCRIPTION
[0033] In order to make the features and advantages of this patent more obvious and easy to understand, the following embodiments are specifically described in detail as follows:
[0034] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.
[0035] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.
[0036] The present invention specifically introduces the algorithm flow through the following embodiments:
[0037] 1 Photovoltaic power prediction process
[0038] The photovoltaic power prediction flow chart based on the BOLL-CNN-BiLSTM deep learning fusion model in the embodiment of the present invention is as follows: Figure 1 As shown, the specific steps are:
[0039] (1) Preprocess the historical data of photovoltaic power and influencing factors. First, remove the data collection points during abnormal periods, and then use the Pearson coefficient to screen the strongly correlated influencing factors to obtain a valid data set after dimensionality reduction;
[0040] (2) Normalize the historical data of PV power and effective influencing factors and divide them into training set and test set;
[0041] (3) Calculate the SMA and BOLL dynamic characteristic parameters of the effective historical data series of photovoltaic power;
[0042] (4) The SMA and BOLL dynamic feature extraction parameters of the calculated historical data are normalized and added into the model input data;
[0043] (5) Input the data sequences of all dimensions into the BOLL-CNN-BiLSTM deep learning model for training and prediction;
[0044] (6) Denormalize the predicted photovoltaic power data and output the results.
[0045] 2 Determination of dynamic feature extraction indicators
[0046] The fluctuation of photovoltaic power generation is the main reason for the uncertainty of photovoltaic power generation. In order to improve the accuracy of photovoltaic power generation prediction, the present invention deeply mines the laws of dynamic data, finds the main factors affecting the fluctuation of photovoltaic power generation, and quantifies the dynamic trends of fluctuation data with different forms, so as to effectively improve the accuracy of photovoltaic power generation prediction in abnormal weather. To this end, two dynamic characteristic parameters are defined, namely Bollinger Bands and Moving Average Lines to analyze the form of fluctuation data to reflect the dynamic characteristics of photovoltaic power generation corresponding to different forms.
[0047] 2.1 Calculation method of moving average indicator
[0048] In time series analysis, the moving average indicator is one of the most popular tools for measuring the overall trend of time series. In actual situations, the simple moving average (SMA) is a traditional moving average indicator. Formula (1) is the calculation formula for the SMA indicator:
[0049]
[0050] Where: p k It represents the photovoltaic power output fluctuation value data of the n days before the i-th day; n represents the number of nodes in the time window scale.
[0051] 2.2 Calculation method of Bollinger Bands indicator
[0052] The Bollinger Bands indicator (BOLL) is a band-shaped technical indicator composed of three tracks. The upper track and the lower track can be regarded as the pressure line and support line of dynamic data respectively. Based on the standard deviation principle in statistics, the channel formed by the two lines specifies the confidence interval for the normal fluctuation of time series data. The middle track can be regarded as the average line of dynamic data, capturing the overall movement trend of time series data. The present invention uses SMA to construct the middle track, eliminate the interference factors in the original sequence, and grasp the overall trend of the original sequence. Formula (2)(3)(4) is the calculation formula of the BOLL indicator:
[0053]
[0054] Where: Indicates the standard deviation of the photovoltaic power output data for the previous n days on the i-th day; BOLL UP , BOLL DN They represent the upper and lower tracks of the Bollinger Bands indicator respectively. α represents the adjustment parameter of the bandwidth range between the upper and lower tracks of the Bollinger Bands indicator. The larger α is, the larger the bandwidth is, and the better the tolerance is to the dynamic changes of the time series.
[0055] 3 Neural Network Model Construction
[0056] 3.1 CNN Convolutional Neural Network
[0057] Convolutional layer and pooling layer are two important components of CNN. Their alternating appearance helps to deeply explore the features of input data. In addition, CNN structure usually also contains a fully connected layer. In order to effectively reduce the dimension of input features, speed up the convergence speed of CNN network and avoid overfitting problems, the pooling layer is usually placed after the convolutional layer. Through convolution and pooling operations, convolutional neural network can learn higher-level feature representation, thereby achieving accurate prediction of complex data. Formula (5) is the specific convolution calculation process formula:
[0058]
[0059] In the formula: K represents the convolution kernel, and its length, width and depth are represented by i, j and n respectively; I represents the input of the neuron; O represents the local output of the neuron; bn represents the threshold of the convolution kernel; represents the matrix multiplication operation; l, m and n represent the three dimensions of the output matrix respectively.
[0060] 3.2 BiLSTM Neural Network
[0061] Recurrent neural networks (RNNs) generally have the problem of vanishing gradients. Although they have been successfully applied to solve time series prediction problems in many fields, during the training process, as the time steps increase, the network's gradient will gradually decrease, resulting in the inability to effectively propagate information from earlier time steps to subsequent time steps. This problem limits the network's ability to model long-term dependencies, also known as the long-term dependency problem. The key to solving this problem is to introduce linear connections and gating units into LSTM neural networks.
[0062] Figure 2 Shows the structure of the LSTM neural network, where h t-1 、h t 、h t+1 Represent the encoded hidden state vectors at different times, and x t-1 、x t 、x t+1 represents the time series input at each moment. In this structure, σ represents the sigmoid function and tanh represents the hyperbolic tangent function. In the LSTM recurrent unit, the hidden layer is mainly composed of four key structures: input gate, output gate, forget gate and cell state. The core of the LSTM recurrent unit is the cell state, and the three gating units that control the flow of the cell state are responsible for the input gate, output gate and forget gate. By cooperating with the cell state, the state information is maintained, updated and transmitted.
[0063] When using traditional unidirectional neural networks to predict photovoltaic power, training is usually carried out from front to back in chronological order. However, this method cannot fully tap the potential information in the photovoltaic power time series, so there are great limitations in the prediction accuracy and data utilization of the model. In contrast, the bidirectional long short-term memory network has the characteristics of mining the intrinsic connection between the photovoltaic power data at the current moment and the past and future data, because it can perform forward propagation and backward propagation at the same time. Therefore, the use of a bidirectional long short-term memory network can better improve the prediction accuracy and data utilization. The structural diagram of the BiLSTM network model is shown in Figure 3 Compared with the traditional unidirectional LSTM neural network structure, the bidirectional recurrent neural network structure of the BiLSTM network can better utilize the local and global information in the time series, thereby improving the model's feature extraction ability and prediction accuracy. Compared with the unidirectional LSTM neural network, this structure overcomes the disadvantage of insufficient data information.
[0064] 3.4 BOLL-CNN-BiLSTM Neural Network
[0065] The fusion model provided by the embodiment of the present invention is mainly composed of a convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) network. The introduction of the CNN model enables the predicted value to better track the actual photovoltaic power generation, and can also reduce the uncertainty of the prediction result through its powerful feature extraction ability. The BiLSTM network can learn long-term dependencies in the sequence and can make up for the shortcomings of the convolutional neural network (CNN) in this regard. It is difficult to extract the temporal correlation and change law of features simultaneously when CNN and BiLSTM are used separately. The SMA and BOLL dynamic feature extraction indicators are introduced to provide more effective influencing factor data input for the prediction model.
[0066] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0067] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0068] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0069] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0070] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0071] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.
[0072] This patent is not limited to the above-mentioned optimal implementation mode. Anyone can derive other forms of photovoltaic power prediction methods based on dynamic feature extraction and CNN-BiLSTM under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of the present invention should be covered by this patent.
Claims
1. A photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM, characterized by: The trend characteristics of photovoltaic power generation are extracted by moving average and Bollinger Band indicators to obtain dynamic feature parameters. The nonlinear local features between multidimensional input data are extracted by fusing CNN and BiLSTM neural networks, and the photovoltaic power generation power prediction results are output after bidirectional feature extraction of time series data.
2. The photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM according to claim 1 is characterized in that: The dataset is constructed as follows: Firstly, the historical data of photovoltaic power and influencing factors are preprocessed, including: removing data collection points during abnormal periods and using the Pearson coefficient to screen strongly correlated influencing factors, to obtain an effective data set after dimensionality reduction; Then the photovoltaic power history data and effective influencing factor data are normalized and divided into training set and test set; Calculate the moving average and Bollinger Bands indicator dynamic characteristic parameters of the effective historical data series of photovoltaic power; According to the calculated historical data moving average and Bollinger Bands indicator dynamic feature extraction parameters, they are normalized and added to the data set; data sequences of all dimensions are input into the fusion CNN and BiLSTM neural network for training and prediction.
3. The photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM according to claim 2 is characterized in that: The fused CNN and BiLSTM neural network denormalizes the predicted photovoltaic power data and outputs the result.
4. The photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM according to claim 1 is characterized in that: Bollinger Bands and moving averages are used to analyze the shape of the fluctuation data to reflect the dynamic characteristics of photovoltaic power generation corresponding to different shapes.
5. The photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM according to claim 1, characterized in that: The calculation formula of the moving average indicator is: Where: p k Represents the photovoltaic power output fluctuation value data for the n days before the i-th day.
6. The photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM according to claim 5 is characterized in that: The middle line of the Bollinger Bands indicator is constructed using a moving average to eliminate interference factors in the original sequence. The calculation is as follows: Where: Indicates the standard deviation of the photovoltaic power output data for the previous n days on the i-th day; BOLL UP , BOLL DN They represent the upper and lower lines of the Bollinger Bands indicator respectively, and α represents the adjustment parameter of the bandwidth range between the upper and lower lines of the Bollinger Bands indicator.
7. The photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM according to claim 1 is characterized in that: The fusion CNN and BiLSTM neural network is composed of a convolutional neural network CNN and a bidirectional long short-term memory network BiLSTM; the CNN model is used to make the predicted value track the actual photovoltaic power generation power, and the BiLSTM network is used to learn the long-term dependency in the sequence.
8. A photovoltaic power prediction system based on dynamic feature extraction and CNN-BiLSTM, characterized in that: include: Dynamic feature parameter extraction module, used to extract photovoltaic power generation trend characteristics through moving average and Bollinger Band indicators to obtain dynamic feature parameters; The prediction module is used to extract nonlinear local features between multi-dimensional input data by fusing CNN and BiLSTM neural networks, and to output photovoltaic power generation prediction results after performing bidirectional feature extraction on time series data.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of a photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM as described in any one of claims 1 to 7 are implemented.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a photovoltaic power prediction method based on dynamic feature extraction and CNN-BiLSTM as described in any one of claims 1 to 7 are implemented.