Ultra-short-time photovoltaic power prediction method and device, electronic equipment and storage medium

Abstract

The invention discloses an ultra-short-time photovoltaic power prediction method and device, electronic equipment and a storage medium, and the method comprises the steps: constructing a to-be-processed radar echo map sequence based on the real-time radar echo map data of a target photovoltaic power station, and constructing a to-be-processed photovoltaic power sequence based on the real-time photovoltaic power generation power of the target photovoltaic power station; inputting the to-be-processed radar echo map sequence and the to-be-processed photovoltaic power sequence into a pre-constructed ultra-short-time photovoltaic power prediction model to obtain a target photovoltaic power prediction sequence; and carrying out reverse normalization processing on the target photovoltaic power prediction sequence, and predicting actual photovoltaic power corresponding to the target photovoltaic power station. The problem that the photovoltaic power of the target photovoltaic power station cannot be accurately predicted due to the influence of the cloud cluster in the prior art is solved, and the effect of accurately predicting the ultra-short-time photovoltaic power of the target photovoltaic power station is achieved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an ultra-short-time photovoltaic power prediction method, apparatus, electronic device, and storage medium. Background Technology

[0002] Ultra-short-time photovoltaic power forecasting (typically referring to forecasts for the next 0-60 minutes) is crucial for real-time grid dispatching, reserve capacity management, ancillary service trading, and the optimized operation of energy storage systems.

[0003] Currently, the challenge in ultra-short-time photovoltaic (UST) power forecasting lies in the sensitivity of photovoltaic output to complex and rapidly changing cloud formations (a key meteorological factor). Whether based on statistical or physical models, the essence lies in inference from historical data or simplified physical processes. However, due to the highly nonlinear and localized dynamic evolution of cloud formations, and the often insufficient sampling frequency and spatial resolution, existing methods struggle to capture the fine spatiotemporal characteristics of clouds and their complex mapping relationship with photovoltaic power. This leads to a significant increase in UST photovoltaic power forecasting errors during periods of drastic weather change (especially rapid cloud cover or relocation), resulting in inaccurate predictions.

[0004] To address the aforementioned issues, improvements are needed in the methods for predicting ultra-short-time photovoltaic power. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for predicting ultra-short-time photovoltaic power, in order to solve the problem in the prior art that the ultra-short-time photovoltaic power of photovoltaic power plants cannot be accurately predicted due to the influence of cloud clusters.

[0006] In a first aspect, embodiments of the present invention provide an ultra-short-time photovoltaic power prediction method, comprising:

[0007] A radar echo sequence to be processed is constructed based on the real-time radar echo data of the target photovoltaic power station, and a photovoltaic power sequence to be processed is constructed based on the real-time photovoltaic power generation of the target photovoltaic power station.

[0008] The radar echo sequence to be processed and the photovoltaic power sequence to be processed are input into a pre-constructed ultra-short time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence.

[0009] The target photovoltaic power prediction sequence is inversely normalized to predict the actual photovoltaic power corresponding to the target photovoltaic power station.

[0010] Secondly, embodiments of the present invention also provide an ultra-short-time photovoltaic power prediction device, comprising:

[0011] The sequence construction module is used to construct a radar echo sequence to be processed based on the real-time radar echo data of the target photovoltaic power station, and to construct a photovoltaic power sequence to be processed based on the real-time photovoltaic power generation of the target photovoltaic power station.

[0012] A photovoltaic power prediction sequence determination module is used to input the radar echo sequence to be processed and the photovoltaic power sequence to be processed into a pre-constructed ultra-short time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence;

[0013] The actual power prediction module is used to perform inverse normalization processing on the target photovoltaic power prediction sequence to predict the actual photovoltaic power corresponding to the target photovoltaic power station.

[0014] Thirdly, embodiments of the present invention also provide an electronic device, comprising:

[0015] At least one processor; and

[0016] A memory communicatively connected to the at least one processor; wherein,

[0017] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the ultra-short-time photovoltaic power prediction method according to any embodiment of the present invention.

[0018] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer instructions, which are used to cause a processor to execute and implement the ultra-short-time photovoltaic power prediction method according to any embodiment of the present invention.

[0019] Fifthly, embodiments of the present invention also provide a computer program product, including a computer program that, when executed by a processor, implements the ultra-short-time photovoltaic power prediction method as described in any of the embodiments of the present invention.

[0020] The technical solution of this invention includes, before inputting the radar echo sequence to be processed and the photovoltaic power sequence to be processed into a pre-constructed ultra-short-time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence, the following steps are taken: pre-constructing an ultra-short-time photovoltaic power prediction model; wherein, the pre-construction of the ultra-short-time photovoltaic power prediction model includes: acquiring at least one training sample and inputting each training sample into an initial deep learning model; wherein, each training sample includes a historical radar echo sequence and a historical photovoltaic power sequence of the target photovoltaic power station; for each training sample, determining the joint coding feature vector corresponding to the current training sample based on the initial deep learning model, and performing future state joint decoding on the joint coding feature to obtain a predicted training sample corresponding to the current training sample; wherein, the predicted training sample includes a future radar potential feature sequence corresponding to the historical radar echo sequence and a future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence; calculating the joint loss of the current training sample and the predicted training sample, and adjusting the parameters of the initial deep learning model based on the joint loss to obtain the ultra-short-time photovoltaic power prediction model. In the pre-construction stage of the ultra-short-time photovoltaic power prediction model in this technical solution, for at least one training sample of the target photovoltaic power plant, the radar echo sequence and photovoltaic power sequence in the training sample are efficiently encoded. On the one hand, a three-layer stacked convolutional long short-term memory network is used to extract deep features from the radar echo image sequence with complex spatial dynamics and time series characteristics. On the other hand, a two-layer stacked long short-term memory network is used to accurately capture the pattern of the photovoltaic power time series. Subsequently, early and deep fusion is performed through global average pooling, feature concatenation, and multi-layer fully connected networks to form a unified joint encoded feature vector containing rich multi-source information. This ensures that the advantages of different modal data are fully utilized and their effective synergistic effect before prediction is achieved, providing high-quality comprehensive information for subsequent joint prediction. Furthermore, the shared joint encoded feature vector is used to drive the ConvLSTM decoder and LSTM encoder in parallel. The ConvLSTM decoder branch is responsible for predicting the potential feature sequence of future radar echo images and capturing the future evolution trend of meteorological conditions, while the LSTM decoder branch is responsible for predicting future ultra-short-time photovoltaic power. This invention addresses the problem in existing technologies where accurate prediction of photovoltaic power generation from target photovoltaic power plants is impossible due to cloud cover. At each time step of the photovoltaic power prediction decoding process, the model incorporates the predicted potential radar features of the corresponding time step as auxiliary input into the LSTM decoder. This establishes an explicit and direct guiding relationship between the future radar state and the photovoltaic power prediction, enabling the photovoltaic power prediction to dynamically reference and respond to the model's understanding of future weather conditions. This significantly improves the accuracy and robustness of the prediction, achieving accurate ultra-short-time photovoltaic power prediction for target photovoltaic power plants. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the content of the embodiments of the present invention and these drawings without creative effort.

[0022] Figure 1 This is a flowchart of an ultra-short-time photovoltaic power prediction method provided in Embodiment 1 of the present invention;

[0023] Figure 2 This is a flowchart of an ultra-short-time photovoltaic power prediction method provided in Embodiment 2 of the present invention;

[0024] Figure 3 This is a schematic diagram of joint encoding of historical radar echo sequence and historical photovoltaic power sequence according to Embodiment 2 of the present invention;

[0025] Figure 4 This is a schematic diagram of decoding a joint coded feature vector according to Embodiment 2 of the present invention;

[0026] Figure 5 This is a schematic diagram of the structure of an ultra-short-time photovoltaic power prediction device according to Embodiment 3 of the present invention;

[0027] Figure 6 This is a schematic diagram of the structure of an electronic device that implements the ultra-short-time photovoltaic power prediction method of this invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention. The acquisition, transmission, storage, use, and processing of data in the technical solutions of this application comply with the relevant provisions of national laws and regulations. It should be noted that in the embodiments of this application, certain software, components, or models and other existing solutions in the industry may be mentioned. These should be considered as exemplary, and their purpose is only to illustrate the feasibility of implementing the technical solutions of this application, but it does not mean that the applicant has or necessarily used such solutions.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.

[0030] Example 1

[0031] Figure 1 The flowchart of an ultra-short-time photovoltaic power prediction method is provided in Embodiment 1 of the present invention. This embodiment is applicable to situations where radar echo data and photovoltaic power data are jointly encoded to mine potential feature information of radar echo data and capture future evolution trends of meteorological conditions, establishing an explicit and direct guiding relationship between the future state of radar and photovoltaic power prediction, and accurately predicting the ultra-short-time photovoltaic power of a target photovoltaic power station. This method can be executed by an ultra-short-time photovoltaic power prediction device, which can be implemented in hardware and / or software and can be configured in a computing device capable of executing the ultra-short-time photovoltaic power prediction method.

[0032] like Figure 1 As shown, the method includes:

[0033] S110. Construct a radar echo sequence to be processed based on the real-time radar echo data of the target photovoltaic power station, and construct a photovoltaic power sequence to be processed based on the real-time photovoltaic power generation of the target photovoltaic power station.

[0034] The target photovoltaic power station refers to a photovoltaic power station that requires ultra-short-time photovoltaic power prediction. Real-time radar echo data refers to continuous spatiotemporal observation data of atmospheric precipitation particle distribution, intensity, and movement trends, detected in real-time by radar and displayed visually in image form. The radar echo sequence to be processed refers to multiple frames of radar echo images corresponding to the stated factual radar echo data, acquired continuously at fixed time intervals. Real-time photovoltaic power generation refers to the electrical power actually generated by the target photovoltaic power station at the current moment, characterizing its instantaneous power generation capacity. The photovoltaic power sequence to be processed refers to a time-series dataset composed of continuous real-time photovoltaic power generation arranged in chronological order, used for subsequent analysis and application.

[0035] In a specific example, radar echo image data of the target photovoltaic power station for one hour (i.e., four time steps) is acquired in real time, with each time step interval fixed at 15 minutes. Each radar echo image is processed into a single-channel two-dimensional image tensor representing the radar echo intensity. Simultaneously, photovoltaic power data is acquired, where the photovoltaic power data corresponds to the time range of the echo image data, such as a sequence of photovoltaic power generation data of the target photovoltaic power station over the past hour (i.e., 12 time steps), with each time step interval fixed at 5 minutes.

[0036] Based on this, the radar echo image data to be processed and the photovoltaic power to be processed are serialized and constructed respectively to obtain the radar echo image sequence to be processed and the photovoltaic power sequence to be processed.

[0037] S120. Input the radar echo sequence and photovoltaic power sequence to be processed into the pre-built ultra-short time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence.

[0038] The ultra-short-time photovoltaic power prediction model is used to predict the ultra-short-time photovoltaic power of a target photovoltaic power plant based on the radar echo sequence to be processed and the photovoltaic power sequence to be processed.

[0039] Based on the above example, the constructed radar echo sequence and photovoltaic power sequence to be processed are precisely time-aligned, and the aligned radar echo sequence and photovoltaic power sequence within the same time range are input into the ultra-short time photovoltaic power prediction model.

[0040] Based on an ultra-short-time photovoltaic power prediction model, spatiotemporal features are extracted from the radar echo sequence to obtain a radar spatiotemporal tensor. Then, global average pooling is applied to the radar spatiotemporal tensor to obtain the radar echo vector to be processed. Simultaneously, based on the ultra-short-time photovoltaic power prediction model, spatiotemporal features are extracted from the photovoltaic power sequence to obtain the photovoltaic power vector to be processed. Further, the radar echo vector and the photovoltaic power vector to be processed are deeply fused and encoded to obtain a joint feature vector.

[0041] Based on this, a dual-head decoder architecture is used to reverse encode the joint feature vector to decode the potential radar features for the next four 15-minute time steps. Then, the photovoltaic power for the next four 15-minute time steps (i.e., the target photovoltaic power prediction sequence) is predicted based on the joint feature vector and the potential radar features.

[0042] S130. Perform inverse normalization on the target photovoltaic power prediction sequence to predict the actual photovoltaic power corresponding to the target photovoltaic power station.

[0043] Based on the above example, the target photovoltaic power prediction sequence output by the ultra-short time photovoltaic power prediction model is precisely denormalized to convert it back to the original megawatt (MW) unit to obtain the actual photovoltaic power.

[0044] In practical applications, actual photovoltaic power is transmitted in real time to the power grid dispatch center, photovoltaic power plant operation and management platform, or other downstream systems via standardized data interfaces. The forecast results are used to guide power grid load dispatch, ancillary service market transactions, optimization of energy storage system charging and discharging strategies, and refined operation and maintenance of photovoltaic power plants. Based on power forecasts for the next 15 to 60 minutes, power grid dispatchers precisely adjust the output of thermal power units or arrange reserve capacity to ensure the safe and stable operation of the power grid.

[0045] The technical solution of this invention involves constructing a radar echo sequence to be processed based on real-time radar echo data of a target photovoltaic power station, and constructing a photovoltaic power sequence to be processed based on the real-time photovoltaic power generation of the target photovoltaic power station. The radar echo sequence and the photovoltaic power sequence to be processed are input into a pre-constructed ultra-short-time photovoltaic power prediction model to obtain a target photovoltaic power prediction sequence. The target photovoltaic power prediction sequence is then subjected to inverse normalization to predict the actual photovoltaic power corresponding to the target photovoltaic power station. In this technical solution, a multi-source heterogeneous spatiotemporal feature encoding and deep fusion mechanism is used to extract spatiotemporal features from the radar echo sequence and photovoltaic power sequence of the target photovoltaic power station, obtaining corresponding vectors to be fused, and then fused and encoded to obtain a joint encoded feature vector. Furthermore, by combining a future state joint prediction and decoding architecture, the intrinsic correlation between cloud dynamics and photovoltaic power changes in the radar echo is deeply explored, thereby enabling the simultaneous prediction of potential features of future radar echoes and future ultra-short-time photovoltaic power. This significantly improves the accuracy of predicting the photovoltaic power of the target photovoltaic power station based solely on radar echo and photovoltaic power data.

[0046] Example 2

[0047] Figure 2This is a flowchart of an ultra-short-time photovoltaic power prediction method provided in Embodiment 2 of the present invention. Optionally, before inputting the radar echo sequence to be processed and the photovoltaic power sequence to be processed into a pre-constructed ultra-short-time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence, the method includes: pre-constructing an ultra-short-time photovoltaic power prediction model; wherein, the pre-construction of the ultra-short-time photovoltaic power prediction model includes: acquiring at least one training sample and inputting each training sample into an initial deep learning model; wherein, each training sample includes a historical radar echo sequence and a historical photovoltaic power sequence of the target photovoltaic power station; for each training sample... The training samples are obtained by determining the joint encoded feature vector corresponding to the current training sample based on the initial deep learning model, and performing future state joint decoding on the joint encoded features to obtain the predicted training sample corresponding to the current training sample; wherein, the predicted training sample includes a future radar potential feature sequence corresponding to the historical radar echo sequence and a future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence; the joint loss of the current training sample and the predicted training sample is calculated, and the parameters of the initial deep learning model are adjusted based on the joint loss to obtain the ultra-short time photovoltaic power prediction model.

[0048] like Figure 2 As shown, the method includes:

[0049] S210. Pre-construct an ultra-short-time photovoltaic power prediction model.

[0050] In practical applications, before inputting the radar echo sequence to be processed and the photovoltaic power sequence to be processed into the pre-built ultra-short time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence, the ultra-short time photovoltaic power prediction model is pre-built.

[0051] Optionally, at least one training sample is obtained and each training sample is input into the initial deep learning model; for each training sample, the joint encoding feature vector corresponding to the current training sample is determined based on the initial deep learning model, and the joint encoding feature is jointly decoded for future states to obtain the predicted training sample corresponding to the current training sample; the joint loss of the current training sample and the predicted training sample is calculated, and the parameters of the initial deep learning model are adjusted based on the joint loss to obtain the ultra-short time photovoltaic power prediction model.

[0052] Each training sample includes a historical radar echo sequence and a historical photovoltaic power sequence of the target photovoltaic power plant. The prediction training samples include a future radar potential feature sequence corresponding to the historical radar echo sequence, and a future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence.

[0053] In this technical solution, each training sample includes 15-minute historical radar echo data and 5-minute historical photovoltaic power generation data, without introducing other external data sources.

[0054] In a specific example, radar echo data for the target photovoltaic power station over the past hour (i.e., four time steps) is retrieved from a historical meteorological database. The interval between each time step is fixed at 15 minutes. Each radar echo mosaic is processed into a single-channel two-dimensional image tensor representing the radar echo intensity.

[0055] Specifically, all radar echo data were cropped and scaled to a uniform spatial resolution of 256 pixels high using bilinear interpolation. A width of 256 pixels ensures precise consistency in the input image size. Based on this, the intensity values ​​are normalized: radar echo intensity values ​​(pixel values ​​in the original range of 0-255) are read and linearly mapped to a floating-point range of 0-1. The specific normalization formula is as follows:

[0056]

[0057] in, Historical radar echo data in coordinates The intensity value at that location, This is the normalized strength. This normalization operation helps accelerate model convergence and improve training stability.

[0058] Furthermore, a sequence of historical radar echo maps corresponding to the historical radar echo map data is constructed based on the following formula:

[0059]

[0060] in , representing the current moment And the normalized radar echo maps of the previous three 15-minute intervals.

[0061] Simultaneously, historical photovoltaic (PV) power data is obtained from a historical meteorological database. This historical PV power data corresponds to the time range of the historical radar echo data; for example, it represents the PV power generation data sequence of the target PV power plant over the past hour (i.e., 12 time steps), with each time step having a fixed interval of 5 minutes. Specifically, the historical PV power data is normalized, mapping it to a floating-point range of 0-1. The specific normalization formula is:

[0062]

[0063] in, This is a normalized value of historical photovoltaic power data. Historical photovoltaic power data, The target photovoltaic power plant's historical maximum photovoltaic power output is fixed at 20 megawatts (20 MW).

[0064] The purpose of this setting is that this normalization operation eliminates the influence of data dimensions, ensuring the accuracy of model learning.

[0065] Based on this, a historical photovoltaic power sequence corresponding to historical photovoltaic power data is constructed:

[0066]

[0067] in, , representing the current moment And the normalized photovoltaic power of the previous eleven 5-minute intervals.

[0068] Furthermore, the existing historical radar echo sequence will be used... and historical photovoltaic power series Accurate time alignment is performed by taking historical radar echo sequence and historical photovoltaic power sequence within the same time range as a set of training samples. The training samples are then fused and encoded based on the initial deep learning model to obtain a joint encoded feature vector. Based on the joint encoded feature vector, a predicted training sample corresponding to the current training sample is obtained.

[0069] Based on the above example, the acquired historical radar echo sequence and historical photovoltaic power series Precise time alignment is required. During model training, input-output sample pairs need to be constructed. Specifically:

[0070] (1) Model input sequence: for the current time step For the prediction task, the model input includes Historical radar echo sequence at the end of the time period and historical photovoltaic power series .

[0071] (2) Definition of Photovoltaic Power Prediction Target: The target of this technical solution is to predict the photovoltaic power values ​​for the next 15, 30, 45, and 60 minutes. Since photovoltaic power data is in 5-minute increments, this corresponds to the 3rd, 6th, 9th, and 12th 5-minute time steps in the future. Therefore, the future photovoltaic power prediction sequence is expressed as:

[0072]

[0073] (3) Definition of radar potential feature prediction targets: In order to achieve joint prediction, it is also necessary to define the future radar potential feature sequence:

[0074]

[0075] Optionally, based on the initial deep learning model, the joint encoded feature vector corresponding to the current training sample is determined, and the joint encoded features are jointly decoded for future states to obtain the predicted training sample corresponding to the current training sample. This includes: extracting spatiotemporal features from historical radar echo sequence based on the first preset encoder to obtain the radar spatiotemporal tensor corresponding to the historical radar echo sequence; extracting spatiotemporal features from historical photovoltaic power sequence based on the second preset encoder to obtain the photovoltaic power time series feature vector corresponding to the historical photovoltaic power sequence; performing fusion encoding based on the radar spatiotemporal tensor and the photovoltaic power time series feature vector to obtain the joint encoded feature vector; and performing parallel decoding of the joint encoded feature vector based on a dual-head decoder architecture to obtain the future radar potential feature sequence corresponding to the historical radar echo sequence and the future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence.

[0076] The first preset encoder is a three-layer stacked convolutional long short-term memory network. The second preset encoder is a two-layer stacked long short-term memory network.

[0077] Specifically, the historical radar echo sequence and historical photovoltaic power sequence are input into the initial deep learning model. For the historical radar echo sequence, the spatiotemporal features are extracted based on the first preset encoder in the initial deep learning model to obtain the radar spatiotemporal tensor.

[0078] In a specific example, such as Figure 3 As shown, a three-layer stacked convolutional long short-term memory (ConvLSTM) network is used as the first preset encoder to extract data from the input radar echo sequence. Spatiotemporal features are extracted from it. Among them, ConvLSTM can simultaneously capture the spatial dependencies of images (through convolution operations) and the temporal dependencies of sequences (through LSTM gating mechanism).

[0079] The operation of each ConvLSTM unit in the ConvLSTM layer structure follows the following gating formula:

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] in, The input feature map for the current time step. and These represent the hidden state and cell state of the previous time step, respectively. This represents the convolution operation. This represents the Hadamard product (element-level multiplication). It is the Sigmoid activation function. The hyperbolic tangent activation function is used. and For learnable convolution kernels and biases.

[0086] In this technical solution, the ConvLSTM encoder configuration includes ConvLSTM layer 1, ConvLSTM layer 2, and ConvLSTM layer 3; wherein, ConvLSTM layer 1 is used to receive signals with a size of... The radar image sequence, wherein the ConvLSTM encoder is configured with 64 filters and the convolution kernel size is [missing information]. The padding method is set to 'same', and return_sequences=True is set to ensure that the output sequence length is the same as the input sequence length. The size of the output feature map of this layer is [size missing]. ConvLSTM layer 2, used to receive the output of ConvLSTM layer 1, is configured with 128 filters and a convolution kernel size of [missing value]. The padding method is set to 'same', and return_sequences=True. The output feature map size of this layer is... ConvLSTM layer 3 receives the output of ConvLSTM layer 2, is configured with 256 filters, and has a convolution kernel size of [missing value]. The padding method is 'same', and return_sequences=False. This layer only outputs the hidden state of the last time step, and the output feature map size is... .

[0087] Based on this, the ConvLSTM encoder ultimately outputs a high-dimensional radar spatiotemporal feature tensor (i.e., the radar spatiotemporal tensor): This tensor can capture detailed spatiotemporal evolution information of the input radar sequence.

[0088] Meanwhile, continue to see Figure 3 A two-layer stacked Long Short-Term Memory (LSTM) network is used as a second preset encoder to obtain historical photovoltaic power timing data from the input. Spatiotemporal features are extracted from historical photovoltaic power sequences. The operation of each LSTM unit in the LSTM layer structure follows the following gating formula:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] in, The input photovoltaic power value at the current time step. and These represent the hidden state and cell state of the previous time step, respectively. and These are the learnable weight matrix and bias vector.

[0095] In this technical solution, the LSTM encoder configuration includes LSTM layer 1 and LSTM layer 2. LSTM layer 1 is used to receive signals with a size of... A photovoltaic power sequence (i.e., the same size as the historical radar echo sequence) is configured with 128 units and return_sequences=True, with the output sequence length being the same as the input sequence length. LSTM layer 2, used to receive the output of LSTM layer 1, is configured with 256 units and return_sequences=False; this layer only outputs the hidden state of the last time step.

[0096] Based on this, the LSTM encoder finally outputs a photovoltaic time-series feature vector containing photovoltaic power sequence time information. (i.e., the photovoltaic power time-series eigenvector).

[0097] Furthermore, based on the radar spatiotemporal tensor and the photovoltaic power time-series feature vector, a fusion encoding is performed to obtain joint encoded features, including: global average pooling of the radar spatiotemporal tensor to obtain the radar spatiotemporal feature vector; and concatenation of the radar spatiotemporal feature vector and the photovoltaic power time-series feature vector to obtain the joint encoded feature vector.

[0098] For details, please refer to [link / reference]. Figure 3 For radar spatiotemporal feature tensor The Global Average Pooling (GAP) operation is applied to transform it into a radar spatiotemporal feature vector. The advantage of this setup is that it effectively compresses the spatial information of radar images while preserving important channel-level features, reducing feature dimensionality, and facilitating fusion with photovoltaic features.

[0099] Based on this, the radar spatiotemporal feature vector With photovoltaic power time series eigenvector Perform a concatenation operation to obtain the joint encoded feature vector:

[0100]

[0101] in, Represents the joint encoded feature vector. Represents the radar spatiotemporal feature vector. This represents the time-series characteristic vector of photovoltaic power.

[0102] Furthermore, composite feature vectors Subsequently, two stacked fully connected layers are used for dimensionality transformation and nonlinear mapping to learn the deep interaction relationships between features of different modalities. Fully connected layer 1 receives 512 input units and outputs 256 units, applying the ReLU activation function; fully connected layer 2 receives 256 input units and outputs 128 units, also applying the ReLU activation function, ultimately outputting a 128-dimensional jointly encoded feature vector. This joint encoded feature vector incorporates rich information from radar charts and photovoltaic power, serving as a common input to the future state joint prediction module.

[0103] The advantage of this setup is that radar echo maps are spatiotemporal data with both spatial and temporal dimensions (two-dimensional image sequences), while photovoltaic power is a one-dimensional time series (numerical sequence). Therefore, this technical solution uses a ConvLSTM encoder to process the radar image. This is because the ConvLSTM encoder can simultaneously capture spatial dependencies (through convolution) and temporal dependencies, and the output hidden state (H_radar) necessarily preserves the spatial structure while encoding temporal information. Therefore, the output radar spatiotemporal feature tensor is a high-dimensional feature tensor (R^(256×256×256)). In contrast, photovoltaic power has no spatial dimension, so using a standard LSTM encoder to directly output a feature vector (H_pv∈R^256) is the most efficient approach. To fuse these two completely different features, global average pooling (GAP) is used on the radar spatiotemporal feature tensor. Its core advantage lies in its ability to efficiently compress the high-dimensional spatiotemporal feature tensor of the radar, while retaining key global spatial summary information (i.e., channel-level features) and converting it into an R^256 feature vector (V_radar). This not only greatly reduces the dimensionality but also enables the radar features to be precisely aligned with the photovoltaic feature vector H_pv in terms of shape, thus creating the necessary conditions for subsequent concatenation fusion.

[0104] Based on this, the joint encoded feature vector is decoded in parallel using a dual-head decoder architecture to obtain the future radar potential feature sequence corresponding to the historical radar echo sequence and the future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence. This includes: decoding the radar spatiotemporal feature vector based on the first decoder corresponding to the first preset encoder to obtain the future radar potential feature sequence corresponding to the historical radar echo sequence; and decoding the photovoltaic power time series feature vector based on the second decoder corresponding to the second preset encoder to obtain the future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence.

[0105] Specifically, a dual-head decoder architecture is used to decode historical radar echo sequences and historical photovoltaic power sequences in parallel, thereby predicting future radar potential feature sequences and future photovoltaic power prediction sequences, thus achieving joint modeling of radar echo changes and photovoltaic power fluctuations. This branch employs a two-layer stacked ConvLSTM decoder to extract jointly encoded features. The potential radar feature sequences for the next four 15-minute time steps were decoded.

[0106] Specifically, such as Figure 4 As shown, the ConvLSTM decoder (i.e., the first decoder) is initialized, and the jointly encoded features are... First, it is mapped to a 4096-dimensional vector through a fully connected layer, and then reshaped to a size of... A three-dimensional tensor is used as the input to the ConvLSTM decoder at the first prediction time step. The inputs for subsequent prediction time steps are generated by the decoder's internal loop. The ConvLSTM decoder configuration includes ConvLSTM decoder layer 1 and ConvLSTM decoder layer 2. ConvLSTM decoder layer 1 receives a 3D tensor of size [size missing]. The feature map is configured with 128 filters and a convolution kernel size of [size missing]. The padding method is set to 'same', and return_sequences=True is set to generate sequence output. The ConvLSTM decoder layer 2 receives the output of ConvLSTM decoder layer 1, is configured with 64 filters, and has a convolution kernel size of [missing information]. The padding method is 'same', and return_sequences=True is set.

[0107] Based on this, the ConvLSTM decoder outputs a feature tensor at each time step, ultimately forming a sequence of potential radar features for the next four time steps:

[0108]

[0109] Each of them These potential radar feature sequences will directly serve as key auxiliary inputs for the photovoltaic power prediction branch, demonstrating the guiding role of future radar echo information in photovoltaic power prediction.

[0110] Meanwhile, continue to see Figure 4 An LSTM decoder (i.e., a second decoder) is used in conjunction with jointly encoded feature vectors. and potential signature sequences for future radars It predicts the future photovoltaic power generation sequence for the next four 15-minute time steps (i.e., 15, 30, 45, and 60 minutes).

[0111] Specifically, LSTM decoder initialization includes the initial hidden state of the LSTM decoder. and initial cell state By joint coding features It is generated through two independent fully connected layers. The layer is used to receive 128 input units and output 256 units to generate... ; The layer is used to receive 128 input units and output 256 units to generate... During the prediction process, the second decoder performs cyclic predictions over four time steps, predicting one photovoltaic power value at each time step. At each prediction time step... (correspond The input of the LSTM decoder It consists of two parts: historical power information and future radar characteristics. The historical power information is: for the first prediction time step (… The input value is 0.0. For subsequent prediction time steps ( Input the normalized photovoltaic power value predicted by the model itself at the previous time step. This strategy avoids dependence on the actual future power and achieves pure autoregressive prediction. Specifically, the future radar features are the potential features of the future radar corresponding to the current time step. A flattening operation is performed to obtain a 4096-dimensional vector, which is then mapped to a 64-dimensional vector through a fully connected layer. This vector represents the future radar echo image information at the corresponding time step. The two parts (1-dimensional historical power information and 64-dimensional future radar features) are concatenated to form a 65-dimensional input vector. LSTM layer: configured with 256 units. At each time step, the LSTM layer receives... And update its hidden state. and cell state Power output: The current hidden state of the LSTM layer. Then, through a fully connected layer (receiving 256 input units and outputting 1 unit) and a Sigmoid activation function, the normalized photovoltaic power prediction value for the current time step is obtained. The Sigmoid function precisely limits the output value to the range of 0-1, consistent with the power normalization range.

[0112] Based on this, the LSTM decoder ultimately outputs a sequence of predicted future photovoltaic power for the next four 15-minute time steps:

[0113]

[0114] Optionally, the joint loss between the current training sample and the predicted training sample is calculated, and the parameters of the initial deep learning model are adjusted based on the joint loss to obtain an ultra-short-time photovoltaic power prediction model. This includes: obtaining the real future radar echo sequence corresponding to the historical radar echo sequence, and the real photovoltaic power sequence corresponding to the historical photovoltaic power sequence; calculating the loss between the real future radar echo sequence and the future radar potential feature sequence based on a pre-set radar potential feature prediction loss function to obtain the radar potential feature prediction loss; calculating the loss between the real photovoltaic power sequence and the future photovoltaic power prediction sequence based on a pre-set photovoltaic power prediction loss function to obtain the photovoltaic power prediction loss; determining the joint loss between the current training sample and the predicted training sample based on the radar potential feature prediction loss and the photovoltaic power prediction loss; and adjusting the parameters of the initial deep learning model based on the joint loss to obtain the ultra-short-time photovoltaic power prediction model.

[0115] Specifically, the joint loss between the current training sample and the predicted training sample is determined based on the radar latent feature prediction loss and the photovoltaic power prediction loss, including: determining a first weight corresponding to the radar latent feature prediction loss and a second weight corresponding to the photovoltaic power prediction loss; obtaining a first value to be used based on the product of the radar latent feature prediction loss and the first weight, and obtaining a second value to be used based on the product of the photovoltaic power prediction loss and the second weight; and obtaining the joint loss between the current training sample and the predicted training sample based on the sum of the first value to be used and the second value to be used.

[0116] In a specific example, the initial deep learning model is trained as a unified end-to-end system, and the parameters of all components are jointly optimized to obtain an ultra-short-time photovoltaic power prediction model, in order to minimize radar latent feature prediction error and photovoltaic power prediction error.

[0117] This technical solution employs a weighted joint loss function. To simultaneously optimize the radar future potential feature prediction branch and the photovoltaic power prediction branch, ensuring that the two tasks complement each other and enabling the model to learn the deep correlation between radar echo changes and photovoltaic power fluctuations.

[0118] Specifically, the mean squared error (MSE) is used to measure the model's predicted sequence of potential radar features for the future. Latent feature sequences extracted from real future radar echo image sequences using the same radar encoder Differences between them:

[0119]

[0120] in, This indicates the loss in radar latent feature prediction. This represents the Frobenius norm, used to calculate the difference between tensors.

[0121] The mean absolute error (MAE) measures the photovoltaic power sequence predicted by the model. Compared with the actual photovoltaic power sequence Differences between them:

[0122]

[0123] in, Indicates the predicted loss of photovoltaic power. This represents the Frobenius norm, used to calculate the difference between tensors.

[0124] Based on this, the joint loss is obtained by considering both radar latent feature prediction loss and photovoltaic power prediction loss:

[0125]

[0126] in, Indicates joint loss, Indicates the first weight. This indicates the loss in radar latent feature prediction. Indicates the second weight. This indicates the predicted loss of photovoltaic power.

[0127] in, It can be fixed as , It can be fixed as Alternatively, custom settings can be made according to the actual situation. The setting of these two coefficients aims to balance the importance of the two loss terms. Photovoltaic power prediction is taken as the main optimization objective, so its loss terms are given higher weight.

[0128] Furthermore, the initial deep learning model is optimized based on the joint loss corresponding to the training samples. Specifically:

[0129] (1) Optimizer selection: The Adam optimizer is used to update the model parameters. The Adam optimizer is selected because it performs well in handling sparse gradients and non-stationary objectives and has a fast convergence speed.

[0130] (2) Learning rate setting: The initial learning rate is set to... To avoid model overfitting and promote convergence, an exponential learning rate decay strategy is adopted, with a decay rate of 0.95, decaying once every 10 training epochs.

[0131] (3) Training batch size: Each training batch is set to 32 samples, which strikes a balance between computational resources and training efficiency.

[0132] (4) Number of training iterations: The model is trained for a total of 100 epochs. After each epoch, the model is evaluated on an independent validation set and the model weights with the smallest loss on the validation set are saved to ensure optimal generalization ability.

[0133] (5) Training process:

[0134] Forward propagation: For each training batch, the input radar echo sequence and photovoltaic power sequence Simultaneously, a joint coded feature is generated through a multi-source heterogeneous spatiotemporal feature encoding and deep fusion module. Subsequently, The input future state is jointly predicted and decoded by the module, and the future radar potential feature sequence is output in parallel. and future photovoltaic power forecast sequence .

[0135] Loss calculation: based on the predicted sequence output by the model. and and the corresponding real target sequence and Calculate joint loss .

[0136] Backpropagation: The gradient is calculated and propagated to all learnable parameters of the model (including weights and biases in ConvLSTM, LSTM, and fully connected layers) through the backpropagation algorithm.

[0137] Parameter Update: The Adam optimizer updates the model parameters based on the calculated gradients to minimize... Thus, an ultra-short-time photovoltaic power prediction model was obtained.

[0138] Understandably, a high-performance computing environment is needed to support model inference operations before practical application. For example, hardware configuration: the deployment environment should be equipped with NVIDIA GPUs, whose computing power must meet the CUDA 11.x compatibility standard to accelerate parallel computation of deep learning models and ensure the timeliness of real-time predictions. Software stack: deploy a Linux operating system and install the Python 3.8.x runtime environment, as well as the TensorFlow 2.10.x deep learning framework, OpenCV 4.x image processing library, NumPy 1.23.x and Pandas 1.5.x data processing libraries. Model loading: accurately load the trained optimal model weight file (HDF5 format) into the pre-configured inference engine to ensure accurate matching of model structure and parameters.

[0139] After the deployment of the ultra-short time photovoltaic power prediction model, it is necessary to access the latest data in real time and perform preprocessing that is strictly consistent with the training phase, including: (1) Real-time radar echo map acquisition: establish a connection with the meteorological radar data service system through a standardized API interface (RESTful API) and acquire the latest radar echo map in real time every 15 minutes. (2) Real-time photovoltaic power data acquisition: establish a connection with the photovoltaic power station real-time data acquisition system through the OPC-UA protocol or Modbus / TCP protocol and acquire the latest photovoltaic power generation data in real time every 5 minutes. The acquired raw power values ​​are immediately normalized in the same way as in step 1.2. (3) Sliding window to construct input sequence: at each prediction time The system automatically collects radar echo sequence data from the past hour (four of the most recent 15-minute radar images). Photovoltaic power sequence over the past hour (12 latest 5-minute photovoltaic power values) These two sequences are continuously updated in the form of a sliding window to ensure the timeliness of the input data.

[0140] S220. Construct a radar echo sequence to be processed based on the real-time radar echo data of the target photovoltaic power station, and construct a photovoltaic power sequence to be processed based on the real-time photovoltaic power generation of the target photovoltaic power station.

[0141] S230. Input the radar echo sequence and photovoltaic power sequence to be processed into the pre-built ultra-short time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence.

[0142] S240. Perform inverse normalization on the target photovoltaic power prediction sequence to predict the actual photovoltaic power corresponding to the target photovoltaic power station.

[0143] The technical solution of this invention includes, before inputting the radar echo sequence to be processed and the photovoltaic power sequence to be processed into a pre-constructed ultra-short-time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence, the following steps are taken: pre-constructing an ultra-short-time photovoltaic power prediction model; wherein, the pre-construction of the ultra-short-time photovoltaic power prediction model includes: acquiring at least one training sample and inputting each training sample into an initial deep learning model; wherein, each training sample includes a historical radar echo sequence and a historical photovoltaic power sequence of the target photovoltaic power station; for each training sample, determining the joint coding feature vector corresponding to the current training sample based on the initial deep learning model, and performing future state joint decoding on the joint coding feature to obtain a predicted training sample corresponding to the current training sample; wherein, the predicted training sample includes a future radar potential feature sequence corresponding to the historical radar echo sequence and a future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence; calculating the joint loss of the current training sample and the predicted training sample, and adjusting the parameters of the initial deep learning model based on the joint loss to obtain the ultra-short-time photovoltaic power prediction model. In the pre-construction stage of the ultra-short-time photovoltaic power prediction model in this technical solution, for at least one training sample of the target photovoltaic power plant, the radar echo sequence and photovoltaic power sequence in the training sample are efficiently encoded. On the one hand, a three-layer stacked convolutional long short-term memory network is used to extract deep features from the radar echo image sequence with complex spatial dynamics and time series characteristics. On the other hand, a two-layer stacked long short-term memory network is used to accurately capture the pattern of the photovoltaic power time series. Subsequently, early and deep fusion is performed through global average pooling, feature concatenation, and multi-layer fully connected networks to form a unified joint encoded feature vector containing rich multi-source information. This ensures that the advantages of different modal data are fully utilized and their effective synergistic effect before prediction is achieved, providing high-quality comprehensive information for subsequent joint prediction. Furthermore, the shared joint encoded feature vector is used to drive the ConvLSTM decoder and LSTM encoder in parallel. The ConvLSTM decoder branch is responsible for predicting the potential feature sequence of future radar echo images and capturing the future evolution trend of meteorological conditions, while the LSTM decoder branch is responsible for predicting future ultra-short-time photovoltaic power. This invention addresses the problem in existing technologies where accurate prediction of photovoltaic power generation from target photovoltaic power plants is impossible due to cloud cover. At each time step of the photovoltaic power prediction decoding process, the model incorporates the predicted potential radar features of the corresponding time step as auxiliary input into the LSTM decoder. This establishes an explicit and direct guiding relationship between the future radar state and the photovoltaic power prediction, enabling the photovoltaic power prediction to dynamically reference and respond to the model's understanding of future weather conditions. This significantly improves the accuracy and robustness of the prediction, achieving accurate ultra-short-time photovoltaic power prediction for target photovoltaic power plants.

[0144] Example 3

[0145] Figure 5 This is a schematic diagram of an ultra-short-time photovoltaic power prediction device provided in Embodiment 3 of the present invention. Figure 5 As shown, the device includes: a sequence construction module 310, a photovoltaic power prediction sequence determination module 320, and an actual power prediction module 330.

[0146] Among them, the sequence construction module 310 is used to construct a radar echo sequence to be processed based on the real-time radar echo data of the target photovoltaic power station, and to construct a photovoltaic power sequence to be processed based on the real-time photovoltaic power generation of the target photovoltaic power station.

[0147] The photovoltaic power prediction sequence determination module 320 is used to input the radar echo image sequence to be processed and the photovoltaic power sequence to be processed into the pre-built ultra-short time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence;

[0148] The actual power prediction module 330 is used to perform inverse normalization processing on the target photovoltaic power prediction sequence to predict the actual photovoltaic power corresponding to the target photovoltaic power station.

[0149] The technical solution of this invention involves constructing a radar echo sequence to be processed based on real-time radar echo data of a target photovoltaic power station, and constructing a photovoltaic power sequence to be processed based on the real-time photovoltaic power generation of the target photovoltaic power station. The radar echo sequence and the photovoltaic power sequence to be processed are input into a pre-constructed ultra-short-time photovoltaic power prediction model to obtain a target photovoltaic power prediction sequence. The target photovoltaic power prediction sequence is then subjected to inverse normalization to predict the actual photovoltaic power corresponding to the target photovoltaic power station. In this technical solution, a multi-source heterogeneous spatiotemporal feature encoding and deep fusion mechanism is used to extract spatiotemporal features from the radar echo sequence and photovoltaic power sequence of the target photovoltaic power station, obtaining corresponding vectors to be fused, and then fused and encoded to obtain a joint encoded feature vector. Furthermore, by combining a future state joint prediction and decoding architecture, the intrinsic correlation between cloud dynamics and photovoltaic power changes in the radar echo is deeply explored, thereby enabling the simultaneous prediction of potential features of future radar echoes and future ultra-short-time photovoltaic power. This significantly improves the accuracy of predicting the photovoltaic power of the target photovoltaic power station based solely on radar echo and photovoltaic power data.

[0150] Optionally, the ultra-short-time photovoltaic power prediction device further includes: a model building module for pre-building an ultra-short-time photovoltaic power prediction model; wherein,

[0151] The model building module includes:

[0152] The sample acquisition submodule is used to acquire at least one training sample and input each training sample into the initial deep learning model; wherein, each training sample includes the historical radar echo sequence and historical photovoltaic power sequence of the target photovoltaic power station;

[0153] The prediction training sample determination submodule is used to determine the joint encoded feature vector corresponding to the current training sample based on the initial deep learning model for each training sample, and to perform joint decoding of the joint encoded features for future states to obtain the prediction training sample corresponding to the current training sample; wherein, the prediction training sample includes the future radar potential feature sequence corresponding to the historical radar echo map sequence, and the future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence.

[0154] The model determination submodule is used to calculate the joint loss of the current training sample and the predicted training sample, and to adjust the parameters of the initial deep learning model based on the joint loss to obtain the ultra-short time photovoltaic power prediction model.

[0155] Optionally, the prediction training sample determination submodule includes: a radar spatiotemporal tensor determination unit, used to extract spatiotemporal features from the historical radar echo sequence based on the first preset encoder, and obtain the radar spatiotemporal tensor corresponding to the historical radar echo sequence; wherein, the first preset encoder is a three-layer stacked convolutional long short-term memory network.

[0156] The photovoltaic power time series feature vector determination unit is used to extract the spatiotemporal features in the historical photovoltaic power sequence based on the second preset encoder to obtain the photovoltaic power time series feature vector corresponding to the historical photovoltaic power sequence; wherein, the second preset encoder is a long short-term memory network based on two stacked layers;

[0157] The joint coding feature vector determination unit is used to perform fusion coding based on radar spatiotemporal tensor and photovoltaic power time series feature vector to obtain joint coding feature vector;

[0158] The decoding unit is used to perform parallel decoding of the joint encoded feature vector based on a dual-head decoder architecture to obtain the potential feature sequence of future radar corresponding to the historical radar echo map sequence, and the prediction sequence of future photovoltaic power corresponding to the historical photovoltaic power sequence.

[0159] Optionally, the joint coding feature vector determination unit includes: a radar spatiotemporal feature vector determination subunit, which is used to perform global average pooling on the radar spatiotemporal tensor to obtain the radar spatiotemporal feature vector;

[0160] The joint coding feature vector determination sub-unit is used to concatenate the radar spatiotemporal feature vector and the photovoltaic power time series feature vector to obtain the joint coding feature vector.

[0161] Optionally, the decoding unit includes: a first subunit, used to decode the radar spatiotemporal feature vector based on a first decoder corresponding to the first preset encoder, to obtain a future radar potential feature sequence corresponding to the historical radar echo sequence;

[0162] The second subunit is used to decode the photovoltaic power time-series feature vector based on the second decoder corresponding to the second preset encoder, so as to obtain the future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence.

[0163] Optionally, the model determination submodule includes: an acquisition unit, used to acquire the real future radar echo sequence corresponding to the historical radar echo sequence, and the real photovoltaic power sequence corresponding to the historical photovoltaic power sequence;

[0164] The first loss determination unit is used to calculate the loss of the real future radar echo sequence and the future radar potential feature sequence based on the pre-set radar potential feature prediction loss function, and obtain the radar potential feature prediction loss.

[0165] The second loss determination unit is used to calculate the loss of the real photovoltaic power sequence and the future photovoltaic power prediction sequence based on a pre-set photovoltaic power prediction loss function, so as to obtain the photovoltaic power prediction loss.

[0166] The joint loss determination unit is used to determine the joint loss between the current training sample and the predicted training sample based on the radar latent feature prediction loss and the photovoltaic power prediction loss.

[0167] The model determination unit is used to adjust the parameters of the initial deep learning model based on the joint loss to obtain the ultra-short-time photovoltaic power prediction model.

[0168] Optionally, the joint loss determination unit includes: a weight determination subunit, used to determine a first weight corresponding to the radar potential feature prediction loss and a second weight corresponding to the photovoltaic power prediction loss;

[0169] The sub-unit for determining the value to be used is used to obtain the first value to be used by multiplying the loss predicted based on radar potential features and the first weight, and to obtain the second value to be used by multiplying the loss predicted based on photovoltaic power and the second weight.

[0170] The joint loss determination subunit is used to obtain the joint loss between the current training sample and the predicted training sample based on the sum of the first and second values ​​to be used.

[0171] The ultra-short-time photovoltaic power prediction device provided in the embodiments of the present invention can execute the ultra-short-time photovoltaic power prediction method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method.

[0172] Example 4

[0173] Figure 6 A schematic diagram of the structure of an electronic device 10 according to an embodiment of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0174] like Figure 6 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0175] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0176] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as ultra-short-time photovoltaic power prediction methods.

[0177] In some embodiments, the ultra-short-time photovoltaic power prediction method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the ultra-short-time photovoltaic power prediction method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the ultra-short-time photovoltaic power prediction method by any other suitable means (e.g., by means of firmware).

[0178] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0179] Computer programs for implementing the ultra-short-time photovoltaic power prediction method of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0180] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0181] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0182] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0183] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0184] Example 5

[0185] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the ultra-short-time photovoltaic power prediction method as provided in any embodiment of this application.

[0186] In implementing the computer program product, computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0187] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0188] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for predicting ultra-short-time photovoltaic power, characterized in that, include: A radar echo sequence to be processed is constructed based on the real-time radar echo data of the target photovoltaic power station, and a photovoltaic power sequence to be processed is constructed based on the real-time photovoltaic power generation of the target photovoltaic power station. The radar echo sequence to be processed and the photovoltaic power sequence to be processed are input into a pre-constructed ultra-short time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence. The target photovoltaic power prediction sequence is inversely normalized to predict the actual photovoltaic power corresponding to the target photovoltaic power station.

2. The method according to claim 1, characterized in that, Before inputting the radar echo sequence and the photovoltaic power sequence to be processed into a pre-built ultra-short-time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence, the process includes: A pre-constructed ultra-short-time photovoltaic power prediction model is used; among which, The pre-built ultra-short-time photovoltaic power prediction model includes: At least one training sample is obtained, and each training sample is input into the initial deep learning model; wherein, each training sample includes the historical radar echo sequence and historical photovoltaic power sequence of the target photovoltaic power station; For each training sample, a joint coding feature vector corresponding to the current training sample is determined based on the initial deep learning model, and the joint coding feature vector is jointly decoded for future states to obtain a predicted training sample corresponding to the current training sample; wherein, the predicted training sample includes a future radar potential feature sequence corresponding to the historical radar echo sequence and a future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence. Calculate the joint loss of the current training sample and the predicted training sample, and adjust the parameters of the initial deep learning model based on the joint loss to obtain the ultra-short-time photovoltaic power prediction model.

3. The method according to claim 2, characterized in that, The step of determining the joint encoded feature vector corresponding to the current training sample based on the initial deep learning model, and performing future state joint decoding on the joint encoded features to obtain the predicted training sample corresponding to the current training sample, includes: Based on the first preset encoder, the spatiotemporal features in the historical radar echo sequence are extracted to obtain the radar spatiotemporal tensor corresponding to the historical radar echo sequence; wherein, the first preset encoder is a three-layer stacked convolutional long short-term memory network. The spatiotemporal features in the historical photovoltaic power sequence are extracted based on the second preset encoder to obtain the photovoltaic power time series feature vector corresponding to the historical photovoltaic power sequence; wherein, the second preset encoder is a long short-term memory network based on two stacked layers; Based on the radar spatiotemporal tensor and the photovoltaic power time-series feature vector, a joint encoded feature vector is obtained by fusion encoding. The joint encoded feature vector is decoded in parallel using a dual-head decoder architecture to obtain a future radar potential feature sequence corresponding to the historical radar echo sequence, and a future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence.

4. The method according to claim 3, characterized in that, The fusion encoding based on the radar spatiotemporal tensor and the photovoltaic power time-series feature vector yields joint encoded features, including: Global average pooling is performed on the radar spatiotemporal tensor to obtain the radar spatiotemporal feature vector; The radar spatiotemporal feature vector and the photovoltaic power time-series feature vector are concatenated to obtain a joint encoded feature vector.

5. The method according to claim 4, characterized in that, The parallel decoding of the jointly encoded feature vector based on the dual-head decoder architecture to obtain the future radar potential feature sequence corresponding to the historical radar echo sequence and the future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence includes: Based on the first decoder corresponding to the first preset encoder, the radar spatiotemporal feature vector is decoded to obtain the potential feature sequence of future radar corresponding to the historical radar echo map sequence. The photovoltaic power time-series feature vector is decoded based on the second decoder corresponding to the second preset encoder to obtain the future photovoltaic power prediction sequence corresponding to the historical photovoltaic power sequence.

6. The method according to claim 2, characterized in that, The step of calculating the joint loss of the current training sample and the predicted training sample, and adjusting the parameters of the initial deep learning model based on the joint loss to obtain the ultra-short-time photovoltaic power prediction model, includes: Obtain the real future radar echo sequence corresponding to the historical radar echo sequence, and the real photovoltaic power sequence corresponding to the historical photovoltaic power sequence; Based on a pre-set radar potential feature prediction loss function, the loss is calculated on the real future radar echo sequence and the future radar potential feature sequence to obtain the radar potential feature prediction loss. Based on a pre-set photovoltaic power prediction loss function, the loss is calculated for the actual photovoltaic power sequence and the future photovoltaic power prediction sequence to obtain the photovoltaic power prediction loss. Based on the radar latent feature prediction loss and the photovoltaic power prediction loss, determine the joint loss between the current training sample and the predicted training sample; Based on the joint loss, the parameters of the initial deep learning model are adjusted to obtain an ultra-short-time photovoltaic power prediction model.

7. The method according to claim 6, characterized in that, The determination of the joint loss between the current training sample and the predicted training sample based on the radar latent feature prediction loss and the photovoltaic power prediction loss includes: Determine a first weight corresponding to the radar potential feature prediction loss and a second weight corresponding to the photovoltaic power prediction loss; A first value to be used is obtained by multiplying the radar potential feature prediction loss and the first weight, and a second value to be used is obtained by multiplying the photovoltaic power prediction loss and the second weight. Based on the sum of the first value to be used and the second value to be used, the joint loss between the current training sample and the predicted training sample is obtained.

8. An ultra-short-time photovoltaic power prediction device, characterized in that, include: The sequence construction module is used to construct a radar echo sequence to be processed based on the real-time radar echo data of the target photovoltaic power station, and to construct a photovoltaic power sequence to be processed based on the real-time photovoltaic power generation of the target photovoltaic power station. A photovoltaic power prediction sequence determination module is used to input the radar echo sequence to be processed and the photovoltaic power sequence to be processed into a pre-constructed ultra-short time photovoltaic power prediction model to obtain the target photovoltaic power prediction sequence; The actual power prediction module is used to perform inverse normalization processing on the target photovoltaic power prediction sequence to predict the actual photovoltaic power corresponding to the target photovoltaic power station.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the ultra-short-time photovoltaic power prediction method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the ultra-short-time photovoltaic power prediction method according to any one of claims 1-7.

Images