WO2018076475A1 - 一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法 - Google Patents
一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法 Download PDFInfo
- Publication number
- WO2018076475A1 WO2018076475A1 PCT/CN2016/109121 CN2016109121W WO2018076475A1 WO 2018076475 A1 WO2018076475 A1 WO 2018076475A1 CN 2016109121 W CN2016109121 W CN 2016109121W WO 2018076475 A1 WO2018076475 A1 WO 2018076475A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- layer
- rbm
- model
- dnn
- accelerated degradation
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16Z—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS, NOT OTHERWISE PROVIDED FOR
- G16Z99/00—Subject matter not provided for in other main groups of this subclass
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Definitions
- the invention relates to the life prediction of a photovoltaic component, in particular to a photovoltaic component accelerated degradation model and a photovoltaic component life prediction method based on a deep learning method.
- the object of the present invention is to provide a photovoltaic module accelerated degradation model and a photovoltaic module life prediction method based on a deep learning method, which constructs a deep neural network DNN by limiting the Boltzmann machine RBM to different acceleration stress conditions (T i , H i , Ra i ) and the corresponding pseudo-failure distribution quantile function Q i (p) is the input vector, use the CD fast learning algorithm to train RBM, seek the model optimal parameter set ⁇ * , and then use the layer-by-layer greedy method DNN trains to build a photovoltaic module accelerated degradation model to predict the expected working life and pseudo-failure life distribution of PV modules under normal stress conditions.
- a photovoltaic module accelerated degradation model and a photovoltaic module life prediction method based on a deep learning method comprising the following steps:
- DNN Constructing a deep neural network DNN by limiting the Boltzmann machine RBM, inputting initial data, DNN is trained by layer-by-layer greedy method to build a photovoltaic module accelerated degradation model, and then predict the expected working life of PV modules under normal stress conditions.
- the specific process of constructing the deep neural network DNN by using the restricted Boltzmann machine RBM in step (2) is as follows: the deep neural network DNN is composed of a restricted Boltzmann RBM superposition, which includes an input layer, multiple hidden layers and outputs.
- the probability model of the layer, the output of the lower layer RBM is used as the input of the upper layer RBM, and the upper and lower layers RBM are connected by the inter-layer weight coefficient to realize the extraction and transmission of the underlying data probability feature to the top layer output.
- the specific process is: setting the DNN input layer to v, If the hidden layer is h and the number of hidden layers is k, then the joint probability distribution P of the DNN model is expressed as:
- m is the number of neural nodes of the hidden layer h in the RBM unit layer.
- the DNN neural network design in step (2) is a total of 6 layers, including an input layer, an implicit layer and an output layer, wherein the hidden layer has a total of 4 layers, each layer includes 100 Nerve node.
- step of using the layer-by-layer greedy method to train the DNN in the step (2) of the present invention specifically includes the following steps:
- the RBM model is trained layer by layer, that is, the initial data is first input: learning training vector different acceleration stress level combination S i , pseudo-failure life distribution quantile function Q i (p), using contrast divergence CD
- the fast learning algorithm trains to obtain the model weight coefficient W 1 of the first layer RBM hidden layer; the first layer RBM hidden layer h 1 is used as the second layer RBM hidden layer input layer, and the second layer model weight coefficient is trained.
- W 2 recursively, until the DNN model output layer weight coefficient W k is obtained ;
- the initial data S i , Q i (p) is used as the supervised data.
- the supervised learning training is used to further finely adjust the parameter values of the DNN model to achieve parameter optimization and obtain accelerated degradation of the photovoltaic module.
- the model thereby extrapolating the pseudo-failure life distribution quantile function Q 0 (p) under normal stress conditions, thereby obtaining the expected working life of the photovoltaic module.
- the use of supervised learning training to further fine-tune the parameters of the DNN model is preferably achieved by the traditional BP algorithm.
- the ATLAS SEC2100 test chamber and the halm-cetis PV photovoltaic module simulator test system are preferably used in the accelerated degradation experiment in step (1).
- the temperature T i is preferably in the range of 41 to 85 ° C
- the humidity H i is in the range of 62 to 85%
- the optical radiation Ra i is in the range of 840 to 1200 W/m 2 .
- the photovoltaic module used in the accelerated degradation experiment in step (1) is preferably a 18W low-power Mono-Si monocrystalline silicon photovoltaic module, each component is packaged by 4 cell sheets, and is divided into 5 blocks/group for accelerated degradation. Test, each group of samples test time is 1000h, taken every 100h into the halm-cetisPV photovoltaic module simulator test system according to IEC61215-2005 STC output power test.
- the present invention constructs a deep neural network DNN by limiting the Boltzmann machine RBM, and differently accelerating stress conditions (T i , H i , Ra i ) and corresponding pseudo-failure life distributions.
- the number function Q i (p) is the input vector
- the RBM is trained by the CD fast learning algorithm
- the model optimal parameter set ⁇ * is sought
- the DNN is trained by the layer-by-layer greedy method to construct the accelerated degradation model of the photovoltaic module, and then predict the normal stress.
- the expected working life and pseudo-failure life distribution of PV modules under conditions are examples of the expected working life and pseudo-failure life distribution of PV modules under conditions.
- Embodiment 1 is a research idea of accelerated degradation modeling of photovoltaic modules based on deep learning prediction in Embodiment 1-2;
- Embodiment 2 is a schematic diagram showing the principle of deep learning prediction theory modeling in Embodiment 1-2;
- FIG. 3 is a schematic structural diagram of an RBM model in Embodiment 2;
- FIG. 4 is a schematic diagram showing the principle of a CD fast learning algorithm in Embodiment 2;
- Embodiment 5 is a process diagram of a reconstruction error algorithm in Embodiment 2.
- FIG. 6 is a schematic diagram of a DNN construction process and a model in Embodiment 2;
- FIG. 7 is a schematic diagram of a DNN learning training process in Embodiment 2.
- Embodiment 8 is a hardware platform for accelerated degradation test of a photovoltaic module in Embodiment 2;
- Figure 10 is a graph showing the results of EL test of some samples to be tested in Example 2.
- FIG. 11 is a flowchart of an ADM module prediction program in Embodiment 2;
- Figure 13 is a graph showing the results of DNN prediction under normal stress in Example 2.
- the pseudo-failure life is obtained according to the output power P di
- the pseudo-failure life distribution quantile function Q i (p) is obtained according to the pseudo-failure life.
- the conventional method in the field can be used, and the doctoral thesis "high reliability long-life product reliability technology can also be referred to. Research, Deng Aimin, 2006.
- the specific process of constructing the deep neural network DNN by using the restricted Boltzmann machine RBM in step (2) is as follows: the deep neural network DNN is a superimposed Boltzmann machine RBM superposition, a probability model containing multiple hidden layers, the lower RBM The output is used as the input of the upper layer RBM, and the upper and lower layers RBM are connected through the inter-layer weight parameter to realize the extraction and transmission of the underlying data probability feature to the top-level output.
- the specific process is as follows: set the DNN input layer to v, the hidden layer to h, and hide If the number of layers is k, then the joint probability distribution P of the DNN model is expressed as:
- m represents the number of nodes of the hidden layer h in the RBM unit layer.
- the step (2) of using the layer-by-layer greedy method to train the DNN includes the following steps:
- the RBM model is trained layer by layer, that is, the initial data is first input: learning training vector different acceleration stress level combination S i , pseudo-failure life distribution quantile function Q i (p), using contrast divergence CD
- the fast learning algorithm trains to obtain the model weight coefficient W 1 of the first layer RBM hidden layer; the first layer RBM hidden layer h 1 is used as the second layer RBM hidden layer input layer, and the second layer model weight coefficient is trained.
- W 2 recursively, until the DNN model output layer weight coefficient W k is obtained ;
- the initial data S i , Q i (p) is used as the supervised data.
- the supervised learning training is used to further finely adjust the parameter values of the DNN model to achieve parameter optimization and obtain accelerated degradation of the photovoltaic module.
- the model thereby extrapolating the pseudo-failure life distribution quantile function Q 0 (p) under normal stress conditions, yields the expected working life of the photovoltaic module.
- the use of supervised learning training to further fine-tune the parameters of the DNN model is preferably achieved by the traditional BP algorithm.
- the accelerated degradation test system for obtaining performance degradation data of the present invention includes a full spectrum weathering comprehensive test chamber ATLAS SEC2100 (available from ATLAS, USA, integrated BBA level MHG solar simulator, which is one of the few globally available component-level photovoltaic products).
- the multi-stress comprehensive test chamber for the simulation of the working environment the solar module simulator test system halm-cetisPV (providing AAA-level transient optical radiation pulses with a stability of up to ⁇ 5W/m 2 for the STC standard test conditions for photovoltaic modules: Key test equipment such as 1000W/m 2 , 25 ⁇ 1°C, AM1.5 output power Pd measurement).
- Figure 8 is a block diagram of the accelerated degradation test platform of the present invention.
- Table 1 shows the key parameters of the ATLAS SEC2100 test chamber for accelerated degradation.
- the test chamber integrates the MHG solar simulator, high and low temperature alternating environment chamber, simulates the comprehensive stress of T i , humidity H i and optical radiation Ra i to achieve the test.
- the accelerated degradation process of the sample is the MHG solar simulator, high and low temperature alternating environment chamber, simulates the comprehensive stress of T i , humidity H i and optical radiation Ra i to achieve the test.
- the comprehensive test chamber can provide a full-spectrum radiation range of 280 nm to 3000 nm, and the optical radiation intensity can be adjusted between 800 W/m 2 and 1200 W/m 2 , and the temperature and humidity ranges cover the component acceleration test parameter setting interval. , fully meet the requirements of accelerated degradation test conditions.
- the halm-cetisPV is composed of AAA transient solar simulator and IV tester.
- the transient simulator can provide the light source matching degree ⁇ 25% and uniformity ⁇ 2.
- test sample is installed in a closed test room of 8m ⁇ 4m ⁇ 3m.
- the inner wall of the test room is blackened to reduce stray light such as background light and reflected light. Impact, and the independent component temperature monitoring device is added during the test to ensure the accuracy of the test results.
- the Mono-Si monocrystalline silicon module was divided into 5 groups/groups for accelerated degradation test.
- the test time of each group of samples was 1000h, and the PV module simulator test system was taken out every 100 hours to perform the STC output power test according to IEC61215-2005.
- the sequence of each test is shown in Table 3.
- Figure 9 is a flow chart of the accelerated degradation test of photovoltaic modules.
- the EL test will be used to inspect the test sample, and 25 test samples with good initial conditions will be screened out.
- EL Electrode, electroluminescence or electroluminescence
- EL Electroluminescence
- the stream is continuously combined to emit light, and the photons are emitted.
- the composite photons are captured by a high-definition CCD camera.
- the captured results are displayed in the form of images after computer processing.
- the EL test can be used to check internal defects such as cracks, fragments, solder joints, and broken gates of photovoltaic modules.
- Fig. 10(A)-(D) is a graph showing the EL test results of some samples to be tested. It can be seen that the sample (d) has a parabolic crack in the cell in the lower right corner, which should be removed from the sample group to be tested, and the remaining samples. (a)-(c) No problem can be used for testing.
- the 25 Mono-Si monocrystalline silicon modules with the initial state of good condition were divided into 5 blocks/groups according to the test procedure of Fig. 9 for accelerated degradation test.
- Table 4 shows the results of accelerated degradation test results for photovoltaic modules.
- Deep learning prediction simulates the multi-layer learning process of the human brain nervous system by constructing a deep neural network. Without the a priori function hypothesis, a combination of low-level features can be formed to form a more abstract high-level representation to find the data distribution characteristics, that is, the physical quantity with complex causality is After appropriate training, learn the rules and use the rules to predict the unknown trends.
- Figure 1 is a research idea of accelerated degradation modeling of photovoltaic modules based on deep learning prediction.
- the Boltzmann machine RBM to construct Deep Neural Networks (DNN)
- the Contrasive Divergence (CD) fast learning algorithm is used to train the RBM and the training results are evaluated; the input PV module accelerates the degradation of the original data.
- CD Contrasive Divergence
- Deep learning is an in-depth study of Machine Learning (ML) theory, compared to early shallow learning models (without intermediate layers or artificial neural networks with few intermediate implicit nodes), with multiple layers of intermediate learning layers ( >3 ⁇ 5 layers) and got its name. Deep learning emphasizes the construction of learning prediction network topology depth and clearly highlights feature expression learning.
- ML Machine Learning
- the deep learning prediction model is a kind of neural network that applies the deep learning algorithm to the deep topology.
- FIG. 2 is a schematic diagram of the principle of deep learning prediction theory modeling.
- the deep learning prediction is to construct multiple hierarchical models, and the output of the previous level is used as the input of the next level, and the learning algorithm is used to extract the features layer by layer, thereby obtaining the input layer raw data and the output layer result. Implied expression. Therefore, the construction of the intermediate level model and the learning algorithm of the inter-layer feature expression are the key to establishing the deep learning prediction model.
- the middle layer model construction method mainly includes Restrictions Boltzmann Machine (RBM) and Auto-Encoders (AE).
- Boltzmann RBM has strong unsupervised learning ability, can learn complex rules in data, and is suitable for fitting probability distribution, especially combined with contrast divergence CD learning training algorithm, which greatly improves the learning of predictive models. effectiveness.
- the following is a discussion of the construction of a single-layer RBM model.
- the DNN is formed by a single-layer RBM stack, and the RBM is trained layer by layer using the CD fast learning algorithm. Finally, the optimal prediction model parameters and the result output are obtained.
- Restricting Boltzmann machine RBM refers to a two-layer structure model including input layer v and hidden layer h.
- the connection weights between layers are W, and the nodes of v and h are independent of each other.
- the model probability distribution P(v,h) satisfies the Boltzmann distribution.
- Figure 3 is a schematic diagram of the structure of the RBM model.
- n and m be the input layer v and the number of hidden layer h nodes, where v i and h j respectively represent the i-th and j-node states of the input layer v and the hidden layer h, and then a given state of the RBM (v) , h) the energy function E(v,h) is:
- W ij is the connection weight of nodes v i , h j
- a i , b j are the bias of the i-th and j-th nodes.
- Z( ⁇ ) is also called a Partition function.
- E data [ ⁇ ] is the conditional expectation under given training data v, which can be obtained by the conditional distribution of the input layer training set v and the hidden layer h node state;
- E model [ ⁇ ] is the RBM model expectation, the item cannot be Direct derivation is obtained, usually using some sampling methods (such as Gibbs sampling, Metropolis-Hastings sampling, etc.) to obtain approximate solutions, but because of the need for more sampling steps, the training efficiency is not high, and the learning speed is slow.
- Contrasive Divergence (CD) fast learning algorithm has the characteristics of high training efficiency and fast learning speed.
- the main idea is to set the input layer training data v as the initial state of sampling, and calculate the conditional probability formula of input layer v and hidden layer h.
- the h-layer node state the next step is to calculate the h-layer node state reconstruction (reconstruction) input layer v', so that the RBM model parameter expectation value approximate solution can be obtained by reconstructing v'.
- the pseudo code of the CD-k fast learning algorithm is as follows:
- RBM (v 1 , v 2 , ..., v n ; h 1 , h 2 , ..., h m ), training sample data set X, sampling times k;
- the weight coefficients w, a, b of the RBM hidden layer can be obtained, thereby obtaining the optimal parameter set ⁇ * .
- RBM needs to use an evaluation index to evaluate the learning outcomes after training.
- the commonly used RBM model evaluation index is the likelihood of training data. Since the likelihood cannot be directly analyzed by mathematical methods, the RBM model can only be evaluated by the approximation method.
- the reconstruction error algorithm is widely used in the evaluation of RBM model learning effects because of its simple principle and low computational complexity.
- the algorithm takes the training data as the initial state, performs a Gibbs sampling, and calculates the reconstructed samples and original training after sampling. Data error is used as an evaluation index, and Figure 5 is a process diagram of the reconstruction error algorithm.
- the DNN is a superposition of the Boltzmann machine RBM, including the input layer, multiple hidden layers, and the output layer.
- the output of the lower layer RBM is used as the input of the upper layer RBM, and the upper and lower layers RBM are connected by the inter-layer weight parameter.
- 6A and 6B are schematic diagrams of a DNN construction process and a model. Let the DNN input layer, the hidden layer be v, h, and the number of hidden layers be k, then the model joint probability distribution P is expressed as:
- m is the number of neural nodes of the hidden layer h in the RBM unit layer.
- DNN training can use a layer-by-layer greedy algorithm
- Figure 7 is a DNN learning training algorithm process diagram.
- the algorithm can be divided into two steps: layered training and overall tuning.
- the first layer RBM hidden layer model weight coefficient W 1 is obtained ; the first layer RBM hidden layer h 1 is used as the second layer RBM input layer, and the second layer model weight coefficient is trained.
- W 2 recursively, until the DNN model output layer weight coefficient W k is obtained ;
- the initial data S i , Q i (p) is used as the supervised data.
- the supervised learning training is used to further finely adjust the parameter values of the DNN model to achieve parameter optimization. It is implemented by the traditional BP algorithm.
- ADM accelerated degradation modeling module
- the hidden layer neural node uses the sigmoid function
- the output layer neural node uses a linear function
- Table 5 shows the DNN deep neural network parameters.
- DNN training uses a layer-by-layer greedy algorithm, and the training process can be divided into two steps:
- FIG. 12 shows the RBM network reconstruction error curve of each layer. 12(A)-(D) are the first layer RBM reconstruction error, the second layer RBM reconstruction error, the third layer RBM reconstruction error, and the fourth layer RBM reconstruction error, respectively.
- the reconstruction error of the first layer RBM is between 0.2 and 0.5, which is rapidly decreased from the high level, and the later trend is slowed.
- the second and third layers of the RBM reconstruction error are respectively oscillated within the range of 0.05 to 0.25, but with The RBM hierarchical reconstruction is deeper, and the reconstruction error of each layer is gradually reduced.
- the reconstruction error of the fourth layer RBM is basically stable within the range of 0.001 to 0.008, which means that the DNN model learning training process enters an equilibrium state.
- the output prediction result after DNN network training and learning is shown in Fig. 13.
- (A) is the DNN prediction quantile function Q(p)
- (B) is the DNN prediction reliability function.
- Table 6 shows the failure life, characteristic lifetime and median lifetime value of the PV module under normal stress using the DNN network, respectively, and compared with the values provided by the manufacturer. It can be seen that using the DNN network prediction results, the failure life, characteristic life and median life value of the PV modules are 22.1 years, 21.6 years, and 21.1 years, respectively, and the error values provided by the manufacturers are 13.1%, 10.7%, and 9.5%. Meet the engineering prediction accuracy requirements.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Photovoltaic Devices (AREA)
Abstract
一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法,该方法通过限制玻尔兹曼机RBM构建深度神经网络DNN,以不同加速应力条件(T i、H i、R ai)及对应伪失效寿命分布分位数函Q i(p)为输入向量,利用CD快速学习算法训练RBM、DNN,寻求模型最优参数集θ *,构建光伏组件加速退化模型,进而预测正常应力条件下光伏组件预期工作寿命。
Description
本发明涉及光伏组件的寿命预测,具体来说一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法。
传统加速模型多利用经验数据统计、物理分析构建,如Arrhenius模型(温度应力)、Eyring模型(温度应力)、逆幂律模型(电应力)等,而光伏组件性能退化应力因素较多(温度T、湿度H、光辐射Ra等),应用时间较短,且结构较为复杂(融合半导体电子、高分子材料、电气设计等),各种应力作用失效机理不一致,直接建立加速应力与产品寿命之间某种明确函数关系将极为困难。针对光伏组件涉及应力参数多,各种应力同时作用且相互影响,直接建立加速应力与产品寿命之间明确函数关系将极为困难。
发明内容
本发明的目的在于提供一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法,该方法通过限制玻尔兹曼机RBM构建深度神经网络DNN,以不同加速应力条件(Ti、Hi、Rai)及对应伪失效寿命分布分位数函Qi(p)为输入向量,利用CD快速学习算法训练RBM,寻求模型最优参数集θ*,再利用逐层贪心法对DNN进行训练,构建光伏组件加速退化模型,进而预测正常应力条件下光伏组件预期工作寿命及伪失效寿命分布。
本发明的上述目的可通过以下的技术措施来实现:一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法,包括以下步骤:
(一)初始数据的获取:选取光伏组件,设定建立加速退化模型不同加速应力水平组合Si,Si包括温度Ti、湿度Hi和光辐射Rai,进行加速退化试验,获取加速退化条件下的光伏组件输出功率Pdi,根据输出功率Pdi获得伪失效寿命TDi,根据伪失效寿命TDi获得伪失效寿命分布分位数函数Qi(p),将不同加速应力条件Si、伪失效寿命分布分位数函数Qi(p)作为初始数据;
(二)利用限制玻尔兹曼机RBM构建深度神经网络DNN,输入初始数据,
采用逐层贪心法对DNN进行训练,构建光伏组件加速退化模型,进而预测正常应力条件下光伏组件预期工作寿命。
在上述基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法中:
步骤(二)中利用限制玻尔兹曼机RBM构建深度神经网络DNN的具体过程为:深度神经网络DNN是由限制玻尔兹曼机RBM叠加,其包含输入层、多个隐含层和输出层的概率模型,下层RBM的输出作为上层RBM的输入,通过层间权值系数连接上下层RBM,实现底层数据概率特征向顶层输出的抽取和传递,具体过程为:设DNN输入层为v、隐含层为h,隐含层数为k,则该DNN模型联合概率分布P表示为:
P(v,h1,…,hk)=P(v|h1)P(h1|h2)…P(hk-1|hk)
式中P(hk-1|hk)是k层RBM单元层相对于k-1层的条件分布,可表示为:
其中m是RBM单元层中隐含层h的神经节点数。
作为本发明的一种优选的实施方式:步骤(二)中DNN神经网络设计共为6层,包括输入层、隐含层及输出层,其中隐含层共为4层,每层包括100个神经节点。
进一步的,本发明步骤(二)中采用逐层贪心法对DNN进行训练的具体包括以下步骤:
2.1分层训练:
自底层输入层开始,对RBM模型逐层进行训练,即首先输入初始数据:学习训练向量不同加速应力水平组合Si、伪失效寿命分布分位数函数Qi(p),利用对比散度CD快速学习算法训练得到第一层RBM隐含层的模型权值系数W1;将第一层RBM隐含层h1作为第二层RBM隐含层输入层,训练得到第二层模型权值系数W2,依次递归,直至得到DNN模型输出层权值系数Wk;
2.2整体调优:
当所有层训练完后,将初始数据Si,Qi(p)作为监督数据,根据最大似然函数,利用监督学习训练进一步微调整个DNN模型参数值,达到参数最优化,得到光
伏组件加速退化模型,从而外推正常应力条件下的伪失效寿命分布分位数函数Q0(p),进而得到光伏组件预期工作寿命。
其中:
分层训练中:输入学习训练向量应力水平组合Si、伪失效寿命分布分位数函数Qi(p)前,优选将Qi(p)值进行归一化处理,所使用的归一化映射函数为:f:Qi(p)→Q'=(Q-Qmin)/(Qmax-Qmin)。
整体调优中:利用监督学习训练进一步微调整个DNN模型参数值时优选采用传统BP算法来实现。
步骤(一)中进行加速退化实验时优选采用ATLAS SEC2100试验箱以及halm-cetisPV光伏组件模拟器测试系统。
步骤(一)中进行加速退化实验时,温度Ti的范围优选是41~85℃、湿度Hi的范围是62~85%、光辐射Rai的范围是840~1200W/m2。
步骤(一)中进行加速退化实验时采用的光伏组件优选为18W小功率Mono-Si单晶硅光伏组件,每个组件由4片电池片连接封装而成,分为5块/组进行加速退化试验,每组样品试验时间为1000h,每隔100h取出放入halm-cetisPV光伏组件模拟器测试系统依据IEC61215-2005进行STC下输出功率测试。
本发明对比现有技术,有如下优点:本发明通过限制玻尔兹曼机RBM构建深度神经网络DNN,以不同加速应力条件(Ti、Hi、Rai)及对应伪失效寿命分布分位数函Qi(p)为输入向量,利用CD快速学习算法训练RBM,寻求模型最优参数集θ*,再利用逐层贪心法对DNN进行训练,构建光伏组件加速退化模型,进而预测正常应力条件下光伏组件预期工作寿命及伪失效寿命分布。
图1为实施例1-2中基于深度学习预测的光伏组件加速退化建模研究思路;
图2为实施例1-2中深度学习预测理论建模原理示意图;
图3为实施例2中RBM模型结构示意图;
图4为实施例2中CD快速学习算法原理示意图;
图5为实施例2中重构误差算法过程图;
图6为实施例2中DNN构建过程及模型示意图;
图7为实施例2中DNN学习训练过程示意图;
图8为实施例2中光伏组件加速退化试验硬件平台;
图9为实施例2中光伏组件加速退化试验流程图;
图10为实施例2中部分待测样品EL测试结果图;
图11为实施例2中ADM模块预测程序流程图;
图12为实施例2中各层RBM网络重构误差曲线图;
图13为实施例2中正常应力下DNN预测结果图。
实施例1
本实施例提供的基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法,包括以下步骤:
(一)初始数据的获取:选取光伏组件,设定建立加速退化模型不同加速应力水平组合Si(i=1,2,3….n,代表不同的加速应力组合),Si包括温度Ti、湿度Hi和光辐射Rai,进行加速退化实验,获取加速退化条件下的光伏组件的输出功率Pdi,根据输出功率Pdi获得伪失效寿命,根据伪失效寿命获得伪失效寿命分布分位数函数Qi(p),将不同加速应力条件Si、伪失效寿命分布分位数函数Qi(p)作为初始数据;
其中根据输出功率Pdi获得伪失效寿命,根据伪失效寿命获得伪失效寿命分布分位数函数Qi(p)可以采用本领域常规方法,也可以参考博士论文《高可靠长寿命产品可靠性技术研究》,邓爱民,2006。
(二)利用限制玻尔兹曼机RBM构建深度神经网络DNN,输入初始数据,采用逐层贪心法对DNN进行训练,构建光伏组件加速退化模型,进而预测正常应力条件下光伏组件预期工作寿命,如图1-2所示。
步骤(二)中利用限制玻尔兹曼机RBM构建深度神经网络DNN的具体过程为:深度神经网络DNN是由限制玻尔兹曼机RBM叠加,包含多个隐含层的概率模型,下层RBM的输出作为上层RBM的输入,通过层间权值参数连接上下层RBM,实现底层数据概率特征向顶层输出的抽取和传递,具体过程为:设DNN输入层为v、隐含层为h,隐含层数为k,则该DNN模型联合概率分布P表示为:
P(v,h1,…,hk)=P(v|h1)P(h1|h2)…P(hk-1|hk)
式中P(hk-1|hk)是k层RBM单元层相对于k-1层的条件分布,可表示为:
其中m代表是RBM单元层中隐含层h的节点数。
步骤(二)中采用逐层贪心法对DNN进行训练的具体包括以下步骤:
2.1分层训练:
自底层输入层开始,对RBM模型逐层进行训练,即首先输入初始数据:学习训练向量不同加速应力水平组合Si、伪失效寿命分布分位数函数Qi(p),利用对比散度CD快速学习算法训练得到第一层RBM隐含层的模型权值系数W1;将第一层RBM隐含层h1作为第二层RBM隐含层输入层,训练得到第二层模型权值系数W2,依次递归,直至得到DNN模型输出层权值系数Wk;
2.2整体调优:
当所有层训练完后,将初始数据Si,Qi(p)作为监督数据,根据最大似然函数,利用监督学习训练进一步微调整个DNN模型参数值,达到参数最优化,得到光伏组件加速退化模型,从而外推正常应力条件下的伪失效寿命分布分位数函数Q0(p),得到光伏组件预期工作寿命。
分层训练中:输入学习训练向量应力水平组合Si、伪失效寿命分布分位数函数Qi(p)前,优选将Qi(p)值进行归一化处理,所使用的归一化映射函数为:
f:Qi(p)→Q'=(Q-Qmin)/(Qmax-Qmin)。
整体调优中:利用监督学习训练进一步微调整个DNN模型参数值时优选采用传统BP算法来实现。
实施例2
本实施例提供的基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法,包括以下步骤:
(1)初始数据的获取:选取光伏组件,设定建立加速退化模型不同加速应力水平组合Si,Si包括温度Ti、湿度Hi和光辐射Rai,进行加速退化实验,获取加速退化条件下的光伏组件的输出功率,根据输出功率获得伪失效寿命分布分位数函Qi(p),将不同加速应力条件Si、伪失效寿命分布分位数函数Qi(p)作为初始数据;
具体过程如下:
1.1、加速退化试验
本发明用于获取性能退化数据的加速退化试验系统包括全光谱耐候性综合试验箱ATLAS SEC2100(美国ATLAS公司提供,集成BBA级MHG太阳模拟器,是全球为数不多可用于组件级光伏产品全态性工作环境模拟的多应力综合试验箱)、光伏组件模拟器测试系统halm-cetisPV(提供稳定度高达±5W/m2的AAA级瞬态光辐射脉冲,用于实现光伏组件STC标准测试条件:1000W/m2,25±1℃,AM1.5下输出功率Pd测量)等关键试验装备。图8为本发明加速退化试验平台框架图。
1.1.1设备配置
表1为本次加速退化用ATLAS SEC2100试验箱关键参数,该试验箱集成MHG太阳模拟器、高低温交变环境箱,模拟Ti、湿度Hi和光辐射Rai等综合应力作用,实现受试样品的加速退化过程。
表1 ATLAS SEC2100全光谱耐候性综合试验箱关键参数
由表1可以看出,该综合试验箱可提供280nm~3000nm全光谱辐射范围,光辐射强度在800W/m2~1200W/m2之间可调,温湿度范围均覆盖组件加速试验参数设置区间,完全满足加速退化试验条件要求。
采用halm-cetisPV光伏组件模拟器测试系统进行测试,halm-cetisPV由AAA瞬态太阳能模拟器、I-V测试仪等组成,其中瞬态模拟器可提供光源匹配度≤±25%、均匀度≤±2%及稳定度≤±0.5%的AAA级高精度瞬态光辐射脉冲,用于光伏组件输出功率Pd测试所需标准模拟太阳光源,I-V测试仪连接待测组件,可测试记录光伏组件输出功率Pd、峰值电压/电流Vmpp/Impp、短路电压/电流Vsc/Isc及填充因子FF等参数。
考虑到环境光强、温度对组件输出功率Pd测试结果影响,试验将测试样品安装于8m×4m×3m封闭试验房内,该试验房内壁涂黑,可减少背景光、反射光等杂散光影响,且测试过程中增加独立组件温度监控装置,保障测试结果精度。
1.1.2初始数据的获取
由于SEC2100全光谱耐候性老化试验箱内部有效光辐射范围仅为700mm×1500mm,无法同时容纳5块市场上常用规格光伏组件进行加速退化试验。因此,定制一批18W小功率Mono-Si单晶硅光伏组件,由4片电池片连接封装而成,表2为该Mono-Si单晶硅组件样品标称规格参数。
表2 Mono-Si单晶硅组件标称规格参数
将Mono-Si单晶硅组件分为5块/组进行加速退化试验,每组样品试验时间为1000h,每隔100h取出放入光伏组件模拟器测试系统依据IEC61215-2005进行STC下输出功率测试。鉴于尽量减小试验箱加速应力调节的行程误差,各次试验顺序见表3。
表3 加速退化试验次序表
图9为光伏组件加速退化试验流程图。投入试验前,为避免样品自身缺陷导致测试数据失真,将利用EL测试对待测试样品进行检查,筛选出初始状态完好的25块测试样品。EL(Electroluminescence,电致发光或场致发光)测试是通过向晶体硅太阳电池外加正向偏置电压,直流电源向电池注入大量非平衡载流子,太阳电池依靠从扩散区注入的非平衡载流子不断复合发光,放出光子,再利用高清晰度CCD相机捕捉复合光子,捕捉结果计算机处理后以图像形式显示,利用EL测试可检查光伏组件隐裂、碎片、虚焊、断栅等内部缺陷。图10(A)-(D)为部分待测样品EL测试结果图,可以看出,样品(d)在右下角电池片有1条抛物状隐裂,应从待测样品组中剔除,其余样品(a)-(c)无问题可用于试验。
将筛选出初始状态完好的25块Mono-Si单晶硅组件分为5块/组按图9试验流程进行加速退化试验,表4为光伏组件加速退化试验结果数据。
表4 Mono-Si单晶硅组件输出功率加速退化试验数据(W)
根据输出功率获得伪失效寿命TDi,根据伪失效寿命TDi获得伪失效寿命分布分位数函Qi(p),将不同加速应力条件Si、伪失效寿命分布分位数函数Qi(p)作为初始数据。
(二)利用限制玻尔兹曼机RBM构建深度神经网络DNN,输入初始数据,采用逐层贪心法对DNN进行训练,构建光伏组件加速退化模型,进而预测正常应力条件下光伏组件预期工作寿命,如图1-2所示。
2.1基于深度学习预测的光伏组件加速退化模型构建
深度学习预测通过构建深度神经网络模拟人脑神经系统多层学习过程,无需先验函数假设,可通过组合低层特征,形成更加抽象高层表示,来寻找数据分布特征,即将具有复杂因果关系的物理量在经过适当训练学习总结规律,并利用总结出规律来预测未知趋势。
图1为基于深度学习预测的光伏组件加速退化建模研究思路。通过限制玻尔兹曼机RBM构建深度神经网络(Deep Neural Networks,DNN),利用对比散度(Contrastive Divergence,CD)快速学习算法训练RBM,并对训练结果进行评估;输入光伏组件加速退化原始数据,构建光伏组件加速退化模型,预测正常应力条件下光伏组件预期工作寿命及可靠性。
2.1.1深度学习预测建模基本原理
深度学习是机器学习(Machine Learning,ML)理论深入研究延续,相对于早期浅层学习模型(没有中间层或仅有极少中间隐含节点的人工神经网络),因具有多层中间学习层(>3~5层)而得名。深度学习强调构建学习预测网络拓扑结构深度且明确突出特征表达学习,通过逐层特征抽取,将输入数据的关键特征表示由低层逐层变换到更抽象的高层特征空间,在保留关键特征信息的同时有效减少数据中蕴含的无效或干扰信息,降低特征信息维度,提高学习效率。深度学习预测模型是将深度学习算法应用于深度拓扑结构的一类神经网络,具有从少数样本集中学习数据集本质特征的能力,对复杂隐含函数的逼近有很好的效果,因此深度学习预测非常适合用于对光伏组件加速退化建模及伪失效寿命预测。图2为深度学习预测理论建模原理示意图。
由图2可以看出,深度学习预测是构造多个层次模型,且上一层次的输出作为下一层次的输入,利用学习算法,逐层进行特征抽取,从而获得输入层原始数据与输出层结果的隐含表达。因此,中间层次模型的构建、层间特征表达的学习算法是建立深度学习预测模型的关键。
2.1.2基于限制玻尔兹曼机RBM的深度学习预测建模方法
前文指出,深度学习预测建模核心在于中间层模型构建及层间学习训练算法。目前中间层模型构建方法主要包括限制玻尔兹曼机(Restrictions Boltzmann Machine,RBM)、自编码器(Auto-Encoders,AE)。其中玻尔兹曼机RBM因具有强大的无监督学习能力、能够学习数据中复杂规则,适合于拟合概率分布等特点,尤其是结合对比散度CD学习训练算法,极大提高预测模型的学习效率。下面首先讨论单层RBM模型构建,通过单层RBM堆叠形成DNN,利用CD快速学习算法逐层训练RBM,最终得到最优化的预测模型参数及结果输出。
(1)限制玻尔兹曼机RBM模型
限制玻尔兹曼机RBM是指一种包含输入层v、隐含层h的两层结构模型,层间连接权重(即权值系数)为W,v、h层节点间相互独立,没有连接,且模型概率分布P(v,h)满足玻尔兹曼分布。图3为RBM模型结构示意图。
设n、m为输入层v、隐含层h节点数,其中vi、hj分别表示输入层v、隐含层h的第i、j个节点状态,则RBM某一给定状态(v,h)的能量函数E(v,h)为:
式中θ={Wij,ai,bj}是RBM模型参数,Wij为节点vi、hj的连接权重,ai、bj为第i、j个节点的偏置(bias)。
基于Boltzmann分布及式(Ⅰ),可得RBM某一给定状态(v,h)的联合概率分布:
式中Z(θ)亦称为配分函数(Partition function)。
从(v,h)的联合概率分布公式,可得基于输入层v的条件概率:
(2)基于CD快速学习算法的RBM模型训练及评估
为寻求θ*最优解,可采用随机梯度上升法(Stochastic Gradient Ascent,SGA)求解ξ(θ)最大值,其迭代公式为:
式中η>0称为学习率,采用SGA法求解ξ(θ)极值的关键在于得到对数似然函数ξ(θ)对于参数θ梯度:
同理,对数似然函数ξ(θ)对于连接权重W,输入层v、隐含层h偏置参数a,b梯度为:
式中Edata[·]为给定训练数据v下的条件期望,可通过输入层训练集v和隐含层h节点状态的条件分布获得;Emodel[·]为RBM模型期望,该项无法直接推导获得,通常利用一些采样方法(如Gibbs采样、Metropolis-Hastings采样等)获取近似解,但由于需要较多采样步数,使得训练效率不高、学习速度缓慢。
对比散度(Contrastive Divergence,CD)快速学习算法具有训练效率高、学习速度快的特点,主要思路是设置输入层训练数据v为采样初始状态,利用输入层v、隐含层h条件概率公式计算h层节点状态,下一步则通过计算出h层节点状态重构(reconstruction)输入层v’,从而通过重构v’可得到RBM模型参数期望值近似解。图4为CD快速学习算法原理示意图。由于CD快速学习算法仅需极少状态转移次数k(大多时k=1),使得RBM模型的学习效率得到了很大提高,下文为CD-k快速学习算法伪代码。
CD-k快速学习算法伪代码如下:
输入:RBM(v1,v2,…,vn;h1,h2,…,hm),训练样本数据集X,采样次数k;
输出:RBM参数梯度估计Δwij,Δai,Δbj,i=1,2,…,n;j=1,2,…,m;
过程:
1.初始化:Δwij=0,Δai=0,Δbj=0;
2.For all the v∈X do
3.v(0)←v;
4.for s=0,…,k-1,k do
5.for j=1,2,…,m do采样hj
(s)~P(hj|v(s));
6.for i=1,2,…,n do采样vi
(s+1)~P(vi|h(s));
7.for j=1,2,…,m,i=1,2,…,n do
8.Δwij←Δwij+[P(hj=1|v(0))·vi
(0)-P(hj=1|v(k))·vi
(k)]
9.Δai←Δai+[v(0)-vi
(k)]
10.Δbj←Δbj+[P(hj=1|v(0))-P(hj=1|v(k))]
因此,采用CD-k算法,可以得到RBM隐含层的权值系数w、a、b,从而获得最优参数集θ*。
RBM作为无监督学习模型,训练学习后还需采用一种评价指标对学习结果进行评估。目前常用的RBM模型评估指标为对训练数据的似然度,由于该似然度无法通过数学方法直接解析得到,故只能采用近似方法对RBM模型进行评估。
重构误差算法因具有原理简单、计算复杂度低等优点,在RBM模型学习效果评估中得到广泛应用,该算法以训练数据为初始状态,执行一次Gibbs采样,计算采样后重构样本与原训练数据误差作为评价指标,图5为重构误差算法过程图。
(3)深度学习预测模型DNN构建
DNN是由限制玻尔兹曼机RBM叠加,包含输入层、多个隐含层、输出层的概率模型,下层RBM的输出作为上层RBM的输入,通过层间权值参数连接上下层RBM,实现底层数据概率特征向顶层输出的抽取和传递。图6A、6B为DNN构建过程及模型示意图。设DNN输入层、隐含层为v、h,隐含层数为k,则该模型联合概率分布P表示为:
P(v,h1,…,hk)=P(v|h1)P(h1|h2)…P(hk-1|hk) (Ⅵ)
式中P(hk-1|hk)是k层RBM单元层相对于k-1层的条件分布,可表示为:
其中m是RBM单元层中隐含层h的神经节点数。
利用RBM叠加构建DNN,需对DNN进行学习训练获取网络参数。DNN训练可采用逐层贪心算法,图7为DNN学习训练算法过程图。
算法可分为分层训练、整体调优两步。
①分层训练:自底层输入开始,对RBM模型逐层进行训练,即首先输入初始数据:学习训练向量不同加速应力水平组合Si、伪失效寿命分布分位数函数Qi(p),利用对比散度CD快速学习算法训练得到第一层RBM隐含层模型权值系数W1;将第一层RBM隐含层h1作为第二层RBM输入层,训练得到第二层模型权值系数W2,依次递归,直至得到DNN模型输出层权值系数Wk;
②整体调优当所有层训练完后,将初始数据Si,Qi(p)作为监督数据,根据最大似然函数,利用监督学习训练进一步微调整个DNN模型参数值,达到参数最优化,可采用传统BP算法来实现。
具体过程如下:
图11为加速退化建模模块(ADM)流程图,其中学习训练输入向量为应力水平组合Si、伪失效寿命分布分位数函数Qi(p){p∈(0:0.01:1)}(其中p是可靠度值),输入训练前将Qi(p)值进行归一化处理,使用的归一化映射函数为:f:Qi(p)→Q'=(Q-Qmin)/(Qmax-Qmin),输出向量为应力水平S0下为伪失效寿命分布分位数值Q0。
DNN网络共设计为6层,包括输入层vin,隐含层hk(k=1,2,3,4,100个神经节点/层)及输出层vout。隐含层神经节点采用sigmoid函数,输出层神经节点采用线性函数,表5为DNN深度神经网络参数。
表5 DNN网络参数
DNN训练采用逐层贪心算法,训练过程可分为两步:
①分层训练:
自底层开始,输入学习训练向量应力水平组合Si、伪失效寿命分布分位数函数再抽样值Qi(p),利用CD快速学习算法训练得到第一层RBM模型权值系数W1;将第一层RBM隐含层h1作为第二层RBM输入层,训练得到第二层模型权值系数W2,依次递归,直至得到DNN输出层权值系数W4;
②整体调优:
当所有层训练完后,将输入样本数据作为监督数据,根据最大似然函数,利用监督学习训练进一步微调整个DNN模型参数值,达到参数最优化目的,图12为各层RBM网络重构误差曲线图,图12(A)-(D)分别为第一层RBM重构误差、第二层RBM重构误差、第三层RBM重构误差、第四层RBM重构误差。
可以看出,第一层RBM重构误差在0.2~0.5之间,由高位快速下降,后期趋势变缓;第二、三层RBM重构误差主体分别在0.05~0.25区间范围内震荡,但随RBM层次重构的深入,每层重构误差逐渐减小,到第四层RBM重构误差基本稳定在0.001~0.008较小范围内,意味着DNN模型学习训练过程进入平衡状态。DNN网络训练学习后输出预测结果如图13所示,(A)为DNN预测分位数函数Q(p)、(B)为DNN预测可靠度函数。
表6为分别利用DNN网络预测光伏组件正常应力下失效寿命、特征寿命和中位寿命值,并与制造商提供值进行比对。可以看出,利用DNN网络预测结果该批光伏组件失效寿命、特征寿命、中位寿命值分别为22.1年、21.6年、21.1年,相对制造商提供值误差为13.1%、10.7%、9.5%,满足工程预测精度需求。
表6 正常应力下失效寿命特征值预测
其中正常应力条件S0是指温度25℃,湿度60%,Ra=800W/m2。
本发明的实施方式不限于此,在本发明上述基本技术思想前提下,按照本领域的普通技术知识和惯用手段对本发明内容所做出其它多种形式的修改、替换或变更,均落在本发明权利保护范围之内。
Claims (8)
- 一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法,其特征在于包括以下步骤:(一)初始数据的获取:选取光伏组件,设定建立加速退化模型不同加速应力水平组合Si,Si包括温度Ti、湿度Hi和光辐射Rai,进行加速退化试验,获取加速退化条件下的光伏组件输出功率Pdi,根据输出功率Pdi获得伪失效寿命TDi,根据伪失效寿命TDi获得伪失效寿命分布分位数函数Qi(p),将不同加速应力条件Si、伪失效寿命分布分位数函数Qi(p)作为初始数据;(二)利用限制玻尔兹曼机RBM构建深度神经网络DNN,输入初始数据,采用逐层贪心法对DNN进行训练,构建光伏组件加速退化模型,进而预测正常应力条件下光伏组件预期工作寿命。
- 根据权利要求1所述的方法,其特征在于:步骤(二)中利用限制玻尔兹曼机RBM构建深度神经网络DNN的具体过程为:深度神经网络DNN是由限制玻尔兹曼机RBM叠加,其包含输入层、多个隐含层和输出层的概率模型,下层RBM的输出作为上层RBM的输入,通过层间权值系数连接上下层RBM,实现底层数据概率特征向顶层输出的抽取和传递,具体过程为:设DNN输入层为v、隐含层为h,隐含层数为k,则该DNN模型联合概率分布P表示为:P(v,h1,…,hk)=P(v|h1)P(h1|h2)…P(hk-1|hk)式中P(hk-1|hk)是k层RBM单元层相对于k-1层的条件分布,可表示为:其中m是RBM单元层中隐含层h的神经节点数。
- 根据权利要求2所述的方法,其特征在于:步骤(二)中深度神经网络DNN共设计为6层,其包括输入层、隐含层及输出层,其中隐含层共为4层,每层包括100个神经节点。
- 根据权利要求2或3所述的方法,其特征在于:步骤(二)中采用逐层贪心法对DNN进行训练的具体包括以下步骤:2.1分层训练:自底层输入层开始,对RBM模型逐层进行训练,即首先输入初始数据:学 习训练向量不同加速应力水平组合Si、伪失效寿命分布分位数函数Qi(p),利用对比散度CD快速学习算法训练得到第一层RBM隐含层的模型权值系数W1;将第一层RBM隐含层h1作为第二层RBM隐含层输入层,训练得到第二层模型权值系数W2,依次递归,直至得到DNN模型输出层权值系数Wk;2.2整体调优:当所有层训练完后,将初始数据Si,Qi(p)作为监督数据,根据最大似然函数,利用监督学习训练进一步微调整个DNN模型参数值,达到参数最优化,得到光伏组件加速退化模型,从而外推正常应力条件下的伪失效寿命分布分位数函数Q0(p),进而得到光伏组件预期工作寿命。
- 根据权利要求4所述的方法,其特征是:整体调优中:步骤(2.2)中利用监督学习训练进一步微调整个DNN模型参数值时采用传统BP算法来实现。
- 根据权利要求1所述的方法,其特征在于:步骤(一)中进行加速退化实验时采用ATLAS SEC2100试验箱以及halm-cetisPV光伏组件模拟器测试系统。
- 根据权利要求1所述的方法,其特征在于:步骤(一)中进行加速退化实验时,温度Ti的范围是41~85℃、湿度Hi的范围是62~85%、光辐射Rai的范围是840~1200W/m2。
- 根据权利要求1所述的方法,其特征在于:步骤(一)中进行加速退化实验时采用的光伏组件为18W小功率Mono-Si单晶硅光伏组件,每个组件由4片电池片连接封装而成,分为5块/组进行加速退化试验,每组样品试验时间为1000h,每隔100h取出放入halm-cetisPV光伏组件模拟器测试系统依据IEC61215-2005进行STC下输出功率测试。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610954517.3A CN106503461B (zh) | 2016-10-26 | 2016-10-26 | 一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法 |
CN201610954517.3 | 2016-10-26 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2018076475A1 true WO2018076475A1 (zh) | 2018-05-03 |
Family
ID=58322440
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2016/109121 WO2018076475A1 (zh) | 2016-10-26 | 2016-12-09 | 一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN106503461B (zh) |
WO (1) | WO2018076475A1 (zh) |
Cited By (27)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109271705A (zh) * | 2018-09-14 | 2019-01-25 | 湘潭大学 | 一种基于深度学习的机器预测性维护方法 |
CN109657937A (zh) * | 2018-11-30 | 2019-04-19 | 西安电子科技大学 | 一种基于退化数据的产品可靠性评估与寿命预测方法 |
CN109977563A (zh) * | 2019-03-29 | 2019-07-05 | 西安建筑科技大学 | 基于tv-hsmm的海底油气管道剩余寿命预测方法 |
CN110232435A (zh) * | 2019-04-30 | 2019-09-13 | 沈阳化工大学 | 一种自适应深度置信网络滚动轴承故障诊断方法 |
CN110263420A (zh) * | 2019-06-17 | 2019-09-20 | 长安大学 | 一种基于bp神经网络的装载机驱动桥极小子样可靠性评估方法 |
CN110309886A (zh) * | 2019-07-08 | 2019-10-08 | 安徽农业大学 | 基于深度学习的无线传感器高维数据实时异常检测方法 |
CN110414718A (zh) * | 2019-07-04 | 2019-11-05 | 上海工程技术大学 | 一种深度学习下的配电网可靠性指标优化方法 |
CN111191835A (zh) * | 2019-12-27 | 2020-05-22 | 国网辽宁省电力有限公司阜新供电公司 | 基于c-gan迁移学习的ies不完备数据负荷预测方法及系统 |
CN111310387A (zh) * | 2020-02-19 | 2020-06-19 | 电子科技大学 | 一种燃料电池寿命预测方法 |
CN111597755A (zh) * | 2020-04-30 | 2020-08-28 | 北京锐达芯集成电路设计有限责任公司 | 基于门控递归单元神经网络的电子器件剩余寿命预测方法 |
CN111650467A (zh) * | 2020-07-22 | 2020-09-11 | 河南科技大学 | 一种电动拖拉机全寿命电磁兼容测试系统及方法 |
CN111709577A (zh) * | 2020-06-17 | 2020-09-25 | 上海海事大学 | 基于长程相关性gan-lstm的rul预测方法 |
CN111814826A (zh) * | 2020-06-08 | 2020-10-23 | 武汉理工大学 | 退役动力电池余能快速检测评级方法 |
CN111898867A (zh) * | 2020-07-05 | 2020-11-06 | 西北工业大学 | 一种基于深度神经网络的飞机总装生产线产能预测方法 |
CN112287813A (zh) * | 2020-10-27 | 2021-01-29 | 西安科技大学 | 一种基于深度信念网络的铁谱图像智能识别方法 |
CN112800616A (zh) * | 2021-02-05 | 2021-05-14 | 中国人民解放军空军工程大学 | 基于比例加速退化建模的设备剩余寿命自适应预测方法 |
CN112926273A (zh) * | 2021-04-13 | 2021-06-08 | 中国人民解放军火箭军工程大学 | 一种多元退化设备剩余寿命预测方法 |
CN113779891A (zh) * | 2021-09-26 | 2021-12-10 | 中机国能电力工程有限公司 | 一种基于数据深度挖掘的光伏组件设计方法 |
CN114298443A (zh) * | 2022-03-03 | 2022-04-08 | 北京航空航天大学杭州创新研究院 | 基于健康状态指数的工业设备预测性维护方法、装置和电子设备 |
CN114692499A (zh) * | 2022-03-30 | 2022-07-01 | 北京时代民芯科技有限公司 | 一种基于机器学习的集成电路芯片寿命的评估方法与系统 |
CN114818345A (zh) * | 2022-05-05 | 2022-07-29 | 兰州理工大学 | 一种光伏组件剩余寿命预测方法及预测系统 |
CN115130326A (zh) * | 2022-08-03 | 2022-09-30 | 哈尔滨工业大学 | 一种平衡力继电器不确定性最大界自动分配方法 |
CN115879610A (zh) * | 2022-11-30 | 2023-03-31 | 中国南方电网有限责任公司超高压输电公司广州局 | 接触器使用期限预测方法、装置、设备和存储介质 |
CN116070484A (zh) * | 2023-01-31 | 2023-05-05 | 南京林业大学 | 一种基于深度学习的在线实时预测构件状态的方法 |
CN117131784A (zh) * | 2023-10-20 | 2023-11-28 | 北京航空航天大学 | 一种基于全局代理优化的加速退化试验设计方法和装置 |
CN117557010A (zh) * | 2024-01-12 | 2024-02-13 | 中国人民解放军火箭军工程大学 | 随机退化系统中备件数量优化方法、系统、设备及介质 |
CN117610324A (zh) * | 2024-01-24 | 2024-02-27 | 西南科技大学 | 一种基于最小偏差度的加速退化试验优化设计方法 |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107122603A (zh) * | 2017-04-24 | 2017-09-01 | 国网山东省电力公司潍坊供电公司 | 一种微电网中蓄电池储能系统寿命测算方法 |
CN107121926A (zh) * | 2017-05-08 | 2017-09-01 | 广东产品质量监督检验研究院 | 一种基于深度学习的工业机器人可靠性建模方法 |
US11657270B2 (en) | 2017-09-28 | 2023-05-23 | Agency For Science, Technology And Research | Self-assessing deep representational units |
CN108426812B (zh) * | 2018-04-08 | 2020-07-31 | 浙江工业大学 | 一种基于记忆神经网络的pm2.5浓度值预测方法 |
CN108776843A (zh) * | 2018-04-12 | 2018-11-09 | 天津大学 | 基于波尔兹曼机的光伏超短期功率预测方法 |
CN110414026B (zh) * | 2018-04-28 | 2023-08-22 | 东君新能源有限公司 | 一种光伏组件使用寿命的评估方法 |
CN112019159A (zh) * | 2019-05-28 | 2020-12-01 | 汉能移动能源控股集团有限公司 | 光伏组件的室外寿命评估方法和装置 |
CN111426957B (zh) * | 2020-05-19 | 2021-06-11 | 华南理工大学 | 一种模拟车辆工况动力电池荷电状态soc估算优化方法 |
TWI732683B (zh) * | 2020-09-17 | 2021-07-01 | 翁敏航 | 太陽能發電模組缺陷智慧診斷系統與方法 |
TWI732682B (zh) * | 2020-09-17 | 2021-07-01 | 翁敏航 | 太陽能發電模組失效分析系統與方法 |
CN115270529A (zh) * | 2022-09-28 | 2022-11-01 | 华东交通大学 | 融合环境因素及随机初始退化的光伏组件可靠性建模方法 |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101576443A (zh) * | 2009-06-16 | 2009-11-11 | 北京航空航天大学 | 基于灰色rbf神经网络的加速寿命试验寿命预测方法 |
CN102270302A (zh) * | 2011-07-20 | 2011-12-07 | 北京航空航天大学 | 一种基于灰色支持向量机的多应力加速寿命试验预测方法 |
US20160260428A1 (en) * | 2013-11-27 | 2016-09-08 | National Institute Of Information And Communications Technology | Statistical acoustic model adaptation method, acoustic model learning method suitable for statistical acoustic model adaptation, storage medium storing parameters for building deep neural network, and computer program for adapting statistical acoustic model |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105095962B (zh) * | 2015-07-27 | 2017-07-28 | 中国汽车工程研究院股份有限公司 | 一种基于bp人工神经网络的材料动态力学性能预测方法 |
-
2016
- 2016-10-26 CN CN201610954517.3A patent/CN106503461B/zh not_active Expired - Fee Related
- 2016-12-09 WO PCT/CN2016/109121 patent/WO2018076475A1/zh active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101576443A (zh) * | 2009-06-16 | 2009-11-11 | 北京航空航天大学 | 基于灰色rbf神经网络的加速寿命试验寿命预测方法 |
CN102270302A (zh) * | 2011-07-20 | 2011-12-07 | 北京航空航天大学 | 一种基于灰色支持向量机的多应力加速寿命试验预测方法 |
US20160260428A1 (en) * | 2013-11-27 | 2016-09-08 | National Institute Of Information And Communications Technology | Statistical acoustic model adaptation method, acoustic model learning method suitable for statistical acoustic model adaptation, storage medium storing parameters for building deep neural network, and computer program for adapting statistical acoustic model |
Non-Patent Citations (3)
Title |
---|
LIU, GUIXIONG ET AL: "Estimation of Pseudo-Failure Lifetime Distribution for Photovoltaic Modules Based on Accelerated Degradation", OPTICS AND PRECISION ENGINEERING, vol. 24, no. 9, 30 September 2016 (2016-09-30), pages 2183 - 2191, XP055605333, ISSN: 1004-0924X, DOI: 10.3788/OPE.20162409.2183 * |
MAO, YONGHUA ET AL.: "Study on application technology of deep learning", APPLICATION RESEARCH OF COMPUTERS, vol. 33, no. 11, 15 June 2016 (2016-06-15), pages 3201 - 3205 * |
ZHANG, WEI ET AL: "Life-prediction of multi-stress accelerated life testing based on BP algorithm of artificial network", ACTA AERONAUTICA ET ASTRONAUTICA SINICA, vol. 30, no. 9, 30 September 2009 (2009-09-30), pages 1691 - 1696 * |
Cited By (44)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109271705A (zh) * | 2018-09-14 | 2019-01-25 | 湘潭大学 | 一种基于深度学习的机器预测性维护方法 |
CN109657937A (zh) * | 2018-11-30 | 2019-04-19 | 西安电子科技大学 | 一种基于退化数据的产品可靠性评估与寿命预测方法 |
CN109657937B (zh) * | 2018-11-30 | 2022-12-30 | 西安电子科技大学 | 一种基于退化数据的产品可靠性评估与寿命预测方法 |
CN109977563A (zh) * | 2019-03-29 | 2019-07-05 | 西安建筑科技大学 | 基于tv-hsmm的海底油气管道剩余寿命预测方法 |
CN109977563B (zh) * | 2019-03-29 | 2022-09-13 | 西安建筑科技大学 | 基于tv-hsmm的海底油气管道剩余寿命预测方法 |
CN110232435A (zh) * | 2019-04-30 | 2019-09-13 | 沈阳化工大学 | 一种自适应深度置信网络滚动轴承故障诊断方法 |
CN110232435B (zh) * | 2019-04-30 | 2022-10-18 | 沈阳化工大学 | 一种自适应深度置信网络滚动轴承故障诊断方法 |
CN110263420A (zh) * | 2019-06-17 | 2019-09-20 | 长安大学 | 一种基于bp神经网络的装载机驱动桥极小子样可靠性评估方法 |
CN110263420B (zh) * | 2019-06-17 | 2022-09-27 | 长安大学 | 一种基于bp神经网络的装载机驱动桥极小子样可靠性评估方法 |
CN110414718A (zh) * | 2019-07-04 | 2019-11-05 | 上海工程技术大学 | 一种深度学习下的配电网可靠性指标优化方法 |
CN110309886B (zh) * | 2019-07-08 | 2022-09-20 | 安徽农业大学 | 基于深度学习的无线传感器高维数据实时异常检测方法 |
CN110309886A (zh) * | 2019-07-08 | 2019-10-08 | 安徽农业大学 | 基于深度学习的无线传感器高维数据实时异常检测方法 |
CN111191835B (zh) * | 2019-12-27 | 2023-06-23 | 国网辽宁省电力有限公司阜新供电公司 | 基于c-gan迁移学习的ies不完备数据负荷预测方法及系统 |
CN111191835A (zh) * | 2019-12-27 | 2020-05-22 | 国网辽宁省电力有限公司阜新供电公司 | 基于c-gan迁移学习的ies不完备数据负荷预测方法及系统 |
CN111310387A (zh) * | 2020-02-19 | 2020-06-19 | 电子科技大学 | 一种燃料电池寿命预测方法 |
CN111310387B (zh) * | 2020-02-19 | 2023-06-09 | 电子科技大学 | 一种燃料电池寿命预测方法 |
CN111597755A (zh) * | 2020-04-30 | 2020-08-28 | 北京锐达芯集成电路设计有限责任公司 | 基于门控递归单元神经网络的电子器件剩余寿命预测方法 |
CN111814826A (zh) * | 2020-06-08 | 2020-10-23 | 武汉理工大学 | 退役动力电池余能快速检测评级方法 |
CN111814826B (zh) * | 2020-06-08 | 2022-06-03 | 武汉理工大学 | 退役动力电池余能快速检测评级方法 |
CN111709577A (zh) * | 2020-06-17 | 2020-09-25 | 上海海事大学 | 基于长程相关性gan-lstm的rul预测方法 |
CN111709577B (zh) * | 2020-06-17 | 2023-08-15 | 上海海事大学 | 基于长程相关性gan-lstm的rul预测方法 |
CN111898867A (zh) * | 2020-07-05 | 2020-11-06 | 西北工业大学 | 一种基于深度神经网络的飞机总装生产线产能预测方法 |
CN111898867B (zh) * | 2020-07-05 | 2023-04-07 | 西北工业大学 | 一种基于深度神经网络的飞机总装生产线产能预测方法 |
CN111650467A (zh) * | 2020-07-22 | 2020-09-11 | 河南科技大学 | 一种电动拖拉机全寿命电磁兼容测试系统及方法 |
CN112287813A (zh) * | 2020-10-27 | 2021-01-29 | 西安科技大学 | 一种基于深度信念网络的铁谱图像智能识别方法 |
CN112287813B (zh) * | 2020-10-27 | 2023-06-23 | 西安科技大学 | 一种基于深度信念网络的铁谱图像智能识别方法 |
CN112800616A (zh) * | 2021-02-05 | 2021-05-14 | 中国人民解放军空军工程大学 | 基于比例加速退化建模的设备剩余寿命自适应预测方法 |
CN112926273A (zh) * | 2021-04-13 | 2021-06-08 | 中国人民解放军火箭军工程大学 | 一种多元退化设备剩余寿命预测方法 |
CN113779891A (zh) * | 2021-09-26 | 2021-12-10 | 中机国能电力工程有限公司 | 一种基于数据深度挖掘的光伏组件设计方法 |
CN114298443A (zh) * | 2022-03-03 | 2022-04-08 | 北京航空航天大学杭州创新研究院 | 基于健康状态指数的工业设备预测性维护方法、装置和电子设备 |
CN114692499A (zh) * | 2022-03-30 | 2022-07-01 | 北京时代民芯科技有限公司 | 一种基于机器学习的集成电路芯片寿命的评估方法与系统 |
CN114818345B (zh) * | 2022-05-05 | 2023-09-12 | 兰州理工大学 | 一种光伏组件剩余寿命预测方法及预测系统 |
CN114818345A (zh) * | 2022-05-05 | 2022-07-29 | 兰州理工大学 | 一种光伏组件剩余寿命预测方法及预测系统 |
CN115130326A (zh) * | 2022-08-03 | 2022-09-30 | 哈尔滨工业大学 | 一种平衡力继电器不确定性最大界自动分配方法 |
CN115130326B (zh) * | 2022-08-03 | 2024-03-12 | 哈尔滨工业大学 | 一种平衡力继电器不确定性最大界自动分配方法 |
CN115879610A (zh) * | 2022-11-30 | 2023-03-31 | 中国南方电网有限责任公司超高压输电公司广州局 | 接触器使用期限预测方法、装置、设备和存储介质 |
CN115879610B (zh) * | 2022-11-30 | 2024-03-29 | 中国南方电网有限责任公司超高压输电公司广州局 | 接触器使用期限预测方法、装置、设备和存储介质 |
CN116070484A (zh) * | 2023-01-31 | 2023-05-05 | 南京林业大学 | 一种基于深度学习的在线实时预测构件状态的方法 |
CN117131784A (zh) * | 2023-10-20 | 2023-11-28 | 北京航空航天大学 | 一种基于全局代理优化的加速退化试验设计方法和装置 |
CN117131784B (zh) * | 2023-10-20 | 2024-02-23 | 北京航空航天大学 | 一种基于全局代理优化的加速退化试验设计方法和装置 |
CN117557010A (zh) * | 2024-01-12 | 2024-02-13 | 中国人民解放军火箭军工程大学 | 随机退化系统中备件数量优化方法、系统、设备及介质 |
CN117557010B (zh) * | 2024-01-12 | 2024-04-05 | 中国人民解放军火箭军工程大学 | 随机退化系统中备件数量优化方法、系统、设备及介质 |
CN117610324A (zh) * | 2024-01-24 | 2024-02-27 | 西南科技大学 | 一种基于最小偏差度的加速退化试验优化设计方法 |
CN117610324B (zh) * | 2024-01-24 | 2024-04-16 | 西南科技大学 | 一种基于最小偏差度的加速退化试验优化设计方法 |
Also Published As
Publication number | Publication date |
---|---|
CN106503461B (zh) | 2018-12-07 |
CN106503461A (zh) | 2017-03-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2018076475A1 (zh) | 一种基于深度学习法构建的光伏组件加速退化模型及光伏组件寿命预测方法 | |
CN109272156B (zh) | 一种超短期风电功率概率预测方法 | |
Chen et al. | Accurate modeling of photovoltaic modules using a 1-D deep residual network based on IV characteristics | |
CN110175386B (zh) | 变电站电气设备温度预测方法 | |
Ciulla et al. | Forecasting the cell temperature of PV modules with an adaptive system | |
CN108256482A (zh) | 一种基于卷积神经网络进行分布学习的人脸年龄估计方法 | |
CN113688869B (zh) | 一种基于生成对抗网络的光伏数据缺失重构方法 | |
CN111832825B (zh) | 融合长短期记忆网络和极限学习机的风电预测方法及系统 | |
Jing et al. | Lifetime prediction of ultraviolet light-emitting diodes using a long short-term memory recurrent neural network | |
CN113659565B (zh) | 一种新能源电力系统频率态势的在线预测方法 | |
CN113781466A (zh) | 基于多尺度注意力密集网络的太阳能电池板缺陷识别方法 | |
CN112001127A (zh) | 一种igbt结温预测的方法 | |
Xing et al. | Comparative study of reformed neural network based short‐term wind power forecasting models | |
Hou et al. | Classification of defective photovoltaic modules in ImageNet-trained networks using transfer learning | |
Liu et al. | Experiment‐based supervised learning approach toward condition monitoring of PV array mismatch | |
Tang et al. | Automatic defect identification of PV panels with IR images through unmanned aircraft | |
Liu et al. | Short‐term photovoltaic power prediction based on MDCM‐GA‐LSTM model | |
Karakose et al. | A Novel Reconfiguration Method Using Image Processing Based Moving Shadow Detection, Optimization, and Analysis for PV Arrays. | |
CN116565856A (zh) | 一种考虑未知拓扑变化的配电网状态估计方法 | |
Yu et al. | Chaotic Extension Neural Network‐Based Fault Diagnosis Method for Solar Photovoltaic Systems | |
Wheeler et al. | Statistical and domain analytics for informed study protocols | |
Song et al. | Research on the life prediction of light-emitting diode based on neural network | |
Sun et al. | FoSSA Optimization‐Based SVM Classifier for the Recognition of Partial Discharge Patterns in HV Cables | |
Yun et al. | [Retracted] Quality Evaluation and Satisfaction Analysis of Online Learning of College Students Based on Artificial Intelligence | |
Zhang et al. | A Deep Learning-Based Surface Defects Detection and Color Classification Method for Solar Cells |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 16919760 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 16919760 Country of ref document: EP Kind code of ref document: A1 |