Photovoltaic power station prediction system based on SEMD-LSTM
By installing data acquisition units and edge computing devices at photovoltaic power plant sites, combined with the improved SEMD-LSTM algorithm and cloud servers, the problems of high cost and inaccurate prediction in existing photovoltaic power plant prediction systems have been solved, achieving efficient and stable management and optimization of photovoltaic power plants.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-30
- Publication Date
- 2026-04-03
AI Technical Summary
Existing photovoltaic power plant information collection systems are costly and consume a lot of power, and fail to fully consider the comprehensive impact of environmental factors, resulting in insufficient accuracy and comprehensiveness of forecast data, especially when facing multiple meteorological elements, making it difficult to meet actual needs.
A photovoltaic power plant forecasting system based on SEMD-LSTM is adopted. By installing acquisition units on the photovoltaic power plant site, meteorological and electrical data are collected in real time. Edge computing devices are used for data processing, and an improved LSTM algorithm with a self-evaluation mechanism is introduced for forecasting. Cloud servers are used for data management and analysis.
It significantly improves the accuracy and stability of photovoltaic power plant forecasts, enhances operational efficiency and safety, reduces operation and maintenance costs, and enables intelligent management and optimized power plant operation.
Smart Images
Figure CN121787615A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent prediction of power systems, and in particular to a photovoltaic power plant prediction system based on SEMD-LSTM, which belongs to the application of photovoltaic power generation system monitoring and machine learning prediction technology. Background Technology
[0002] Some distributed photovoltaic (PV) power plant information acquisition systems typically consist of a monitoring platform, acquisition terminal equipment, and a communication system. The monitoring platform includes an acquisition server and monitoring workstations, communicating with the acquisition devices via specific power communication protocols. The monitoring workstations provide various human-machine interfaces, including screens, graphs, curves, and reports, allowing users to intuitively monitor and analyze the PV power plant's operation. The communication system usually relies on dedicated lines provided by local operators to ensure stable and reliable data transmission. The acquisition devices are responsible for measuring various environmental parameters, such as voltage, current, and illuminance, then processing the data and uploading it to the monitoring platform via a GPRS wireless module.
[0003] However, existing data acquisition systems have several drawbacks. Firstly, their high cost and power consumption pose significant challenges for large-scale deployment. Secondly, current technologies typically consider only some environmental factors or fail to adequately account for the combined impact of these factors. Furthermore, they lack self-evaluation mechanisms to influence prediction results, leading to insufficient accuracy and comprehensiveness of the forecast data. This is particularly true when dealing with multiple influencing factors such as meteorological elements, where existing technologies often fall short of practical requirements. Summary of the Invention
[0004] To address the aforementioned problems, this invention proposes a photovoltaic power plant prediction system based on SEMD-LSTM to improve the accuracy and comprehensiveness of predictions.
[0005] To achieve the above objectives, the photovoltaic power plant prediction system based on SEMD-LSTM proposed in this invention includes a data acquisition unit, an edge computing device, and a cloud server. The data acquisition unit includes a weather sensor, a current sensor, and a voltage sensor. The data acquisition unit is connected to the edge computing device, which includes a storage module, a computing module, and a connection module. The storage module stores the acquired data received from the data acquisition unit. The computing module analyzes the data stored in the storage module and uses the SEMD-LSTM algorithm to predict photovoltaic power. The connection module communicates with the cloud server, sending the data stored in the storage module and the photovoltaic power prediction data from the computing module to the cloud server. The SEMD-LSTM algorithm includes the following steps: Step 1: Select meteorological, current, and voltage data collected at specific times each day within a specific number of days in the past as the raw data for photovoltaic power prediction; Step 2: Use standardized formulas The input data is preprocessed, whereby... For the standardized output, This is the original data. The mean is calculated based on the original data. The standard deviation of the original data; Step 3: Establish an LSTM network. The LSTM processes the multidimensional input sequence data step by step through the control of the input gate, forget gate, and output gate. First, based on the current input and the previous hidden state The new input gate and forget gate values are calculated to update the memory cell state. (Forget gate...) Determine the memory of the previous moment How much information should be retained in the input gate? and candidate memory units The amount of new information introduced is determined jointly, among which, For the output of the forget gate, for Activation function Here is the weight matrix for the forget gate. Let the input vector be the input vector at the current time. The hidden state weight matrix is... This is the hidden state from the previous moment. Let be the bias vector of the forget gate. This is the activation value of the input gate. Here is the weight matrix of the input gate. This is the weight matrix of the previous hidden state. Let be the bias vector of the input gate. Let tanh represent the candidate cell state, and tanh be the hyperbolic tangent function. This is the weight matrix for the candidate cell states. This is the weight matrix of the previous hidden state. This is the bias vector for the candidate cell state. Next, the updated memory cell state is transmitted via... The process involves merging the data, preserving important historical information, and adding new inputs. This represents the current state of the cell. This represents the cell state at the previous moment.
[0006] Finally, output gate pass By determining new hidden states, the model's memory and output are dynamically adjusted to achieve effective prediction of the input sequence. This is the activation value of the output gate. Here is the weight matrix of the output gate. This is the weight matrix of the previous hidden state. This is the bias vector for the output gate. The current hidden state; Step 4: Introduce a self-evaluation factor feedback mechanism to calculate the error between the model's predicted value and the actual observed value at the same time point. An evaluation factor is generated using the error. The output of the hidden state is adjusted by evaluation factors. Finally, the adjusted hidden state is passed to the output layer to generate the final output. These are actual observed values. This is the predicted value for that time point. For mapping functions, This is the adjustment coefficient; Step 5: Repeat the above four steps to predict the photovoltaic power at the current prediction point using the raw data.
[0007] This invention collects real-time environmental meteorological and electrical data through acquisition units installed in photovoltaic power plants. Data processing is performed using edge computing devices, employing an improved SEMD-LSTM algorithm and incorporating a self-evaluation mechanism. This allows the model to autonomously adjust future predictions based on previous prediction errors, significantly improving prediction accuracy and stability. The predictions fully consider meteorological factors related to the photovoltaic power plant environment, enhancing their comprehensiveness. The collected and predicted data can be viewed via a cloud server, improving the operational efficiency and safety of the photovoltaic power plant, enhancing overall management, and reducing maintenance costs. Attached Figure Description
[0008] Figure 1 This is a schematic diagram of the structure of a photovoltaic power plant prediction system based on SEMD-LSTM; Figure 2 Diagram of ConvLSTM network structure; Figure 3 For stacked ConvLSTM network structures; Figure 4 It is a SEMD-LSTM network structure.
[0009] The present invention will now be described in further detail with reference to the accompanying drawings. Detailed Implementation
[0010] See Figure 1The SEMD-LSTM-based photovoltaic power plant forecasting system comprises a data acquisition unit, an edge computing device, and a cloud server. The data acquisition unit is deployed at the photovoltaic power plant site to collect meteorological and electrical data from the environment. The unit includes meteorological sensors, current sensors, and voltage sensors. The meteorological sensors include some or all of the following: wind speed, wind direction, light intensity, air pressure, temperature, humidity, cloud image monitoring, and rainfall sensors. These sensors are used to collect real-time data on wind speed, wind direction, light intensity, air pressure, temperature, humidity, cloud image, rainfall, current, and voltage at the photovoltaic power plant site. These sensors are positioned at different locations within the photovoltaic power plant to ensure comprehensive coverage and accurate and reliable data collection. The data acquisition unit connects to the edge computing device via a RS-485 bus, transmitting the collected data to the edge computing device.
[0011] Edge computing devices are installed at the photovoltaic power plant site and include storage, computing, and connectivity modules. The storage module is configured with a high-performance local database (such as MySQL or Redis with a data persistence strategy) to store the data received from the acquisition unit. The database is optimized to support high-frequency data write operations. The connectivity module is equipped with a 4G wireless network module or a 4G industrial router to achieve a stable communication connection with the cloud server. This module is configured to support automatic reconnection and failover mechanisms to ensure network connectivity stability and continuity. During data transmission, including the data flow from the acquisition unit to the edge computing device and the data flow from the edge computing device to the cloud server, the AES-256-CFB (Cipher Feedback) encryption algorithm is used to ensure the security of data transmission and storage. AES-256-CFB is a symmetric encryption algorithm with high security and suitability for streaming data encryption, ensuring the security and integrity of data transmission, guaranteeing the timeliness and reliability of photovoltaic power plant data, and providing a solid foundation for subsequent data analysis and decision-making. The computing module performs data analysis on the data stored in the storage module, such as calculating average and peak values, and uses an improved LSTM (Long Short-Term Memory) algorithm—SEMD-LSTM—to predict photovoltaic power.
[0012] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network (RNN) algorithm, particularly well-suited for processing time-series data and capable of capturing long-term dependencies within the data. Traditional LSTMs include components such as the forget gate. The forgetting gate determines the memory of the previous moment. How much information should be retained? Among them, For the output of the forget gate, Here is the weight matrix for the forget gate. For sigmoid activation function, This is the hidden state from the previous moment. For the current input, The bias vector for the forget gate; the input gate. and candidate memory units They jointly decide on the introduction of new information, among which, This is the activation value of the input gate. Here is the weight matrix of the input gate. Let be the bias vector of the input gate. Let tanh represent the candidate cell state, and tanh be the hyperbolic tangent function. This is the weight matrix for the candidate cell states. This is the bias vector for the candidate cell state. The updated memory cell state is obtained through... The process involves merging the data, preserving important historical information, and adding new inputs. This represents the current state of the cell. This represents the cell state at the previous time step. Finally, the output gate... pass Adjust the hidden state and generate the output for the current moment. This is the activation value of the output gate. Here is the weight matrix of the output gate. This is the bias vector for the output gate. This represents the hidden state at the current time step. Through these gates, LSTM can effectively capture long-term dependencies in a sequence and handle complex time series tasks. Since the sampled data in this invention is multi-dimensional, feature extraction is performed on the input data, and a convolutional neural network is introduced to achieve multi-dimensional time series prediction, namely ConvLSTM (Convolutional LSTM), whose network structure is as follows: Figure 2 As shown.
[0013] Because meteorological data in photovoltaic power plant environments exhibit significant spatiotemporal dependence—for example, wind speed, temperature, and humidity not only change over time but also correlate with environmental conditions at different spatial locations—a stacked ConvLSTM algorithm is used to capture these complex spatiotemporal dependencies, forming a model such as... Figure 3The structure shown enhances the model's feature extraction capabilities on complex multidimensional data, making it suitable for more complex time series tasks. However, this method is prone to decreased prediction accuracy when dealing with complex multidimensional data, especially when the model lacks a feedback mechanism for errors between historical data and actual observations. To address this issue, this invention proposes an improved Long Short-Term Memory (LSTM) network algorithm, called SEMD-LSTM (Self-Evaluative Multidimensional LSTM). Its core lies in combining the advantages of stacked ConvLSTM algorithms while introducing a self-evaluation mechanism, enabling the model to autonomously adjust future predictions based on previous prediction errors, thereby significantly improving prediction accuracy and stability. Its structure is as follows: Figure 4 As shown.
[0014] The SEMD-LSTM algorithm includes the following steps.
[0015] Step 1: Select Input Data. Select meteorological, current, and voltage data collected at specific times each day within a specific number of days in the past as the raw data for photovoltaic power prediction. The preferred number of days is 30 days, and the preferred time of day is the hour. The meteorological data is data collected by meteorological sensors, including all or part of the data on wind speed, wind direction, light intensity, air pressure, temperature, humidity, cloud image, and rainfall.
[0016] Step 2: Preprocess the input data. Since the input data contains multiple features with different dimensions and scales, the differences in magnitude of these features may cause the model to tend to focus on features with larger magnitudes during training, while ignoring features with smaller magnitudes. Standardization transforms all features to the same scale, eliminating the influence of different dimensions on model training and allowing the model to treat each feature fairly. A standardization formula is used. The input data is preprocessed, whereby... For the standardized output, This is the original data. The mean is calculated based on the original data. This represents the standard deviation of the original data.
[0017] Step 3: Establish an LSTM network. The LSTM processes the multidimensional input sequence data step by step by controlling the input gate, forget gate, and output gate.
[0018] First, based on the current input and the previous hidden state The new input gate and forget gate values are calculated to update the memory cell state. (Forget gate...) Determine the memory of the previous moment How much information should be retained, of which... For the output of the forget gate, for Activation function Here is the weight matrix for the forget gate. Let the input vector be the input vector at the current time. The hidden state weight matrix is... This is the hidden state from the previous moment. This is the bias vector for the forget gate.
[0019] Input gate and candidate memory units The amount of new information introduced is determined jointly, among which, This is the activation value of the input gate. Here is the weight matrix of the input gate. This is the weight matrix of the previous hidden state. Let be the bias vector of the input gate. Let tanh represent the candidate cell state, and tanh be the hyperbolic tangent function. This is the weight matrix for the candidate cell states. This is the weight matrix of the previous hidden state. is the bias vector for the candidate cell state.
[0020] Next, the updated memory cell state is transmitted via... The process involves merging the data, preserving important historical information, and adding new inputs. This represents the current state of the cell. This represents the cell state at the previous moment.
[0021] Finally, output gate pass By determining new hidden states, the model's memory and output are dynamically adjusted to achieve effective prediction of the input sequence. This is the activation value of the output gate. Here is the weight matrix of the output gate. This is the weight matrix of the previous hidden state. This is the bias vector for the output gate. This represents the current hidden state.
[0022] Step 4: Introduce a self-evaluation factor feedback mechanism. Calculate the error between the model's predicted value and the actual observed value at the same time point. ,in, These are actual observed values. This is the predicted value at that specific time point. Then, this error is used to generate an evaluation factor. ,in, The mapping function is typically chosen as the sigmoid function, and the output of the hidden state is adjusted using an evaluation factor. ,in, This is an adjustment coefficient used to control the evaluation factor. The degree of influence on the adjustment of hidden states is relatively large. This will make the effect of the evaluation factors more obvious, and smaller ones This will make the adjustment smoother. Finally, the adjusted hidden state is passed to the output layer to generate the final output.
[0023] Step 5: Output Prediction. Repeat the above four steps to predict the photovoltaic power at the current prediction point using the raw data.
[0024] By iterating through the above steps, the model can generate a series of photovoltaic power predictions for consecutive time points. This prediction is the model's final prediction of photovoltaic power at the current time point t. This prediction integrates historical input data (such as meteorological and electrical parameters over the past 30 days), the LSTM network's ability to memorize and extract sequential data, and the dynamic adjustment of model performance by the self-evaluation factor feedback mechanism.
[0025] This invention utilizes stacked multi-layer LSTM units, with each layer's updates recursively. For the Nth LSTM layer, the state update formula is: 1. Forgotten Gate: ; 2. Input Gate: ; 3. Cell state renewal: ; 4. Hidden output: .
[0026] From the relationships between the implementations in the structure diagram, it can be easily obtained that: The input unit state of each stage is the output state of the previous stage: ; The input to each level is the hidden output of the previous level: .
[0027] Then, after introducing the self-error correction function, only the hidden input of the first stage needs to be modified: for time ; For the hidden output of the upper level of error correction .
[0028] Adjust the output of each gate in the first stage: The forgetting gate calculates the information to be discarded. :
[0029] Calculate the input gate information at the current time. :
[0030] Calculate the cell state at the current moment. :
[0031] The hidden output at this time is calculated by the output gate. :
[0032] The SEMD-LSTM model, incorporating a self-evaluation mechanism, possesses significant potential advantages, enabling its application in multi-dimensional time series forecasting. First, by evaluating model prediction errors in real time, SEMD-LSTM can dynamically adjust its hidden state output, making the model more adaptable to changing environments or sudden events. Second, this self-adjusting capability allows the model to progressively optimize its predictions, reducing accumulated errors and thus improving long-term forecast accuracy. Furthermore, due to this mechanism, the model can more quickly identify and correct prediction biases, thereby enhancing its robustness and stability. These characteristics enable SEMD-LSTM to demonstrate superior predictive power and reliability when handling complex and highly uncertain multi-dimensional data.
[0033] Using data such as wind speed, wind direction, light intensity, temperature, humidity, air pressure, rainfall, and current and voltage already collected by the acquisition unit, and combining this with real-time data collected by the acquisition unit on the same day, photovoltaic power is predicted.
[0034] The above prediction algorithms are all based on the data already collected by the acquisition unit. For a new environment where the equipment has just started operating, it is usually necessary to collect data for a period of time and accumulate a certain dataset before the prediction function can be realized based on the above data prediction algorithms.
[0035] The cloud server receives, stores, and displays the data collected by the edge computing devices and the predicted power generation data. Administrators can view this data on the cloud server. By monitoring the difference between the predicted results and the actual photovoltaic power, combined with meteorological data, potential severe weather conditions, such as cloudy / rainy days or strong winds, can be identified, allowing for proactive countermeasures. Furthermore, the predicted photovoltaic power can be used to optimize the tilt angle of the photovoltaic panels, adjusting the angle on days with more sunshine to maximize power generation and lowering it on days with less sunshine to reduce losses. Simultaneously, the prediction results can help optimize the layout of the battery packs, improve energy storage efficiency, and be used for load management and scheduling, ensuring appropriate measures are taken when power supply is insufficient. In addition, by analyzing the predicted power generation capacity, system anomalies can be detected promptly, allowing for equipment maintenance and repairs to ensure the normal operation of the photovoltaic system. Finally, the prediction results support economic decision-making, such as optimizing electricity sales strategies during peak electricity price periods to increase economic benefits. Comprehensive utilization of these prediction results can significantly improve the efficiency, reliability, and economy of photovoltaic power generation systems.
[0036] This invention collects real-time environmental meteorological and electrical data through acquisition units installed in photovoltaic power plants. Data processing is performed using edge computing devices, employing an improved SEMD-LSTM algorithm and incorporating a self-evaluation mechanism. This allows the model to autonomously adjust future predictions based on previous prediction errors, significantly improving prediction accuracy and stability. The predictions fully consider meteorological factors of the photovoltaic power plant environment, enhancing their comprehensiveness. Collected and predicted data can be viewed through a cloud server, improving the operational efficiency and safety of the photovoltaic power plant, enhancing overall management, and reducing maintenance costs. This invention, through its three key subsystems—acquisition units, edge computing devices, and cloud servers—achieves full-link coverage from real-time monitoring to intelligent decision-making, providing users with personalized intelligent decision support and enabling intelligent management of photovoltaic power plant operations. This invention enables remote monitoring and diagnosis of photovoltaic power generation equipment, timely detection of equipment faults and anomalies, thereby achieving preventative maintenance, reducing downtime, and improving equipment reliability and stability. Simultaneously, the system stores a large amount of collected data in the cloud, providing power plant managers with detailed data reports and analysis results, helping them make more scientific and effective decisions and optimize power plant operation and management. By achieving energy conservation, emission reduction, and environmental benefits, photovoltaic power stations can utilize solar energy resources more efficiently, reduce energy consumption, and reduce carbon emissions, which aligns with the goals of sustainable development and makes a positive contribution to environmental protection.
Claims
1. A photovoltaic power plant prediction system based on SEMD-LSTM, characterized in that, The system includes a data acquisition unit, an edge computing device, and a cloud server. The data acquisition unit includes a weather sensor, a current sensor, and a voltage sensor. The data acquisition unit is connected to the edge computing device, which includes a storage module, a computing module, and a connection module. The storage module stores the acquired data received from the data acquisition unit. The computing module analyzes the data stored in the storage module and uses the SEMD-LSTM algorithm to predict photovoltaic power. The connection module communicates with the cloud server, sending the data stored in the storage module and the photovoltaic power prediction data from the computing module to the cloud server. The SEMD-LSTM algorithm includes the following steps: Step 1: Select meteorological, current, and voltage data collected at specific times each day within a specific number of days in the past as the raw data for photovoltaic power prediction; Step 2: Use standardized formulas The input data is preprocessed, whereby... For the standardized output, This is the original data. The mean is calculated based on the original data. The standard deviation of the original data; Step 3: Establish an LSTM network. The LSTM processes the multidimensional input sequence data step by step through the control of the input gate, forget gate, and output gate. First, based on the current input and the previous hidden state The new input gate and forget gate values are calculated to update the memory cell state. (Forget gate...) Determines the memory of the previous moment How much information should be retained in the input gate? and candidate memory units The amount of new information introduced is determined jointly, among which, For the output of the forget gate, for Activation function Here is the weight matrix for the forget gate. Let the input vector be the input vector at the current time. The hidden state weight matrix is... This is the hidden state from the previous moment. Let be the bias vector of the forget gate. This is the activation value of the input gate. Here is the weight matrix of the input gate. This is the weight matrix of the previous hidden state. Let be the bias vector of the input gate. Let tanh represent the candidate cell state, and tanh be the hyperbolic tangent function. This is the weight matrix for the candidate cell states. This is the weight matrix of the previous hidden state. This is the bias vector for the candidate cell state. Next, the updated memory cell state is transmitted via... The process involves merging the data, preserving important historical information, and adding new inputs. This represents the current state of the cell. This represents the cell state at the previous moment. Finally, output gate pass By determining new hidden states, the model's memory and output are dynamically adjusted to achieve effective prediction of the input sequence. This is the activation value of the output gate. This is the weight matrix of the output gate. This is the weight matrix of the previous hidden state. This is the bias vector for the output gate. The current hidden state; Step 4: Introduce a self-evaluation factor feedback mechanism to calculate the error between the model's predicted value and the actual observed value at the same time point. An evaluation factor is generated using the error. The output of the hidden state is adjusted by evaluation factors. Finally, the adjusted hidden state is passed to the output layer to generate the final output. These are actual observed values. This is the predicted value for that time point. For mapping functions, This is the adjustment coefficient; Step 5: Repeat the above four steps to predict the photovoltaic power at the current prediction point using the raw data.
2. The photovoltaic power plant prediction system based on SEMD-LSTM according to claim 1, characterized in that, The meteorological sensors include some or all of the following: wind speed sensor, wind direction sensor, light sensor, air pressure sensor, temperature sensor, humidity sensor, cloud image monitoring sensor, and rainfall sensor.