CN110059878B - Photovoltaic power generation power prediction model based on CNN LSTM and construction method thereof - Google Patents

Abstract

The invention belongs to the field of photovoltaic power prediction, and discloses a photovoltaic power generation power prediction model based on CNN LSTM and a construction method; consists of two CNN coiled layers, LSTM, attention mechanism and full connection layer. When only power data exist, feature extraction is carried out on the input one-dimensional power generation time sequence through the double-layer convolution layer, a plurality of data matrixes are obtained through 3 x 1 convolution kernel operation, and information abstraction from the time sequence to the high-level features is realized; the attention mechanism carries out automatic attention on LSTM hidden layer output vectors obtained by each matrix through an LSTM algorithm, and a larger weight is given to characteristic quantities which are obviously related to the current output quantity; and processing the output vector of the attention mechanism into a one-dimensional vector to be input into the full-connection layer through expansion operation, and directly outputting the predicted value of the photovoltaic power generation power at the next moment by the full-connection layer. The method has the advantages of small data quantity, high transportability and high prediction precision.

Description

Photovoltaic power generation power prediction model based on CNN LSTM and construction method thereof

Technical Field

The invention belongs to the technical field of prediction or optimization, and particularly relates to a photovoltaic power generation power prediction model based on CNN LSTM and a construction method thereof.

Background

Currently, the closest prior art:

common photovoltaic power generation power prediction methods mainly include a physical method, a regression method, a time series method and a machine learning method. The physical method uses geographic information, photovoltaic component parameters, solar irradiance, atmospheric temperature and other parameters to construct a physical prediction model, and relies on meteorological parameters of numerical weather forecast (NWP) to predict photovoltaic power generation power. Japanese scholars propose a physical prediction model for predicting the electrical energy output of a photovoltaic array using solar radiation intensity as input; a simple physical prediction model based on observation and satellite remote sensing inversion radiation data, which is proposed by Mayer and the like, is applied to a grid-connected photovoltaic power station of a Germany Munich trade exhibition center, the generalization capability of models constructed by two physical methods is relatively poor, errors are obviously increased when the models constructed by the two physical methods meet different weather conditions, the errors are changed because the physical models have fewer parameters and poorer flexibility, and the average error of the models constructed by the physical methods is generally 10-25% under various weather conditions, so the physical method has a general prediction effect. The model constructed by the physical method is expressed by a physical formula, so that the model is easy to understand and explain, but the physical formula needs to be optimized differently aiming at different photovoltaic power stations. Hassanzadeh at the university of Nevada proposes an improved ARMA model, and generates a model and a prediction by using the generated energy data of each hour in sunny days generated by the independent photovoltaic system of the NV energy company headquarter roof 75k W, wherein the prediction error of the result is between 23.0% and 43.0%, so that the prediction capability of a time sequence is poor, and the error fluctuation is large. The regression method, the time series method and the machine learning method are data driving methods, modeling prediction is carried out on the basis of photovoltaic power generation historical data or NWP historical meteorological parameters, and compared with the regression method and the time series method, the machine learning method is a more intelligent prediction method, so that the generalization capability of prediction can be effectively improved, the fluctuation of prediction results caused by weather changes can be reduced, and errors can be reduced. There are a number of applications, among which BP neural networks and SVMs are commonly used algorithms in machine learning.

The deep learning is the latest and popular research field of artificial intelligence and is primarily applied to the field of photovoltaic power generation prediction, and the greatest difference between the deep learning and the traditional machine learning method is that the deep learning can automatically learn the representation of features from data and extract deep features of the data, the deep features can contain thousands of parameters, and the parameters are often not interpretable. The LSTM network is an improved time-cycle neural network (RNN), which is improved after being proposed, and an additional forgetting gate is added. The improved LSTM network solves the problem of gradient disappearance in model training, can learn long-term and short-term dependence information of a time sequence, is the most successful RNN framework at present, and is applied to a plurality of scenes. The basic unit of the LSTM network comprises a forgetting gate, an input gate and an output gate. Forget to input X in door_tAnd a state memory cell S_t-1Intermediate output h_t-1Jointly determine the forgetting part of the state memory unit. X in input Gate_tAnd determining the retention vector in the state memory unit together after the sigmoid and the tanh function are changed respectively. Intermediate output h_tFrom updated S_tAnd an output o_tCo-determined as shown in formula (1) -formula (6):

f_t＝σ(W_fxx_t+W_fhh_t-1+b_f) (1)

i_t＝σ(W_ixx_t+W_ihh_t-1+b_i) (2)

gt＝φ(W_gxx_t+W_ghh_t-1+b_g) (3)

o_t＝σ(W_oxx_t+W_ohh_t-1+b_o) (4)

S_t＝g_t·i_t+S_t-1·f_t (5)

h_t＝o_t·φ(S_t) (6)

in the formula f_t，i_t，g_t，o_t，h_tAnd S_tRespectively the states of the forgetting gate, the input node, the output gate, the intermediate output and the state unit. W_fx，W_fh，W_ix，W_ih，W_gx，W_gh，W_oxAnd W_ohRespectively corresponding gate and input x_tAnd an intermediate output h_t-1A multiplied matrix weight; b_f，b_i，b_g，b_oRespectively being offset terms of corresponding gates(ii) a Representing the bitwise multiplication of elements in a vector; sigma represents sigmoid function variation; phi denotes the change in the tanh function.

The attention mechanism is a neural network simulating the brain attention form, namely more attention is allocated to key things and less attention is allocated to other things at a specific moment so as to achieve the purpose of reasonably utilizing computing resources. The attention mechanism is applied to the deep neural network, so that the neural network adaptively screens out the features which are obviously related to the current output in the input vector, and the interference of other features is reduced, thereby obviously improving the generalization performance of the model. Hidden layer output vector H ═ { H) by LSTM₁,h₂,...h_tAs the input of the attention mechanism, the attention mechanism will look for the characteristic quantity h_iAttention weight parameter α of_iThis is obtained from formulae (7) and (8):

e_i＝tanh(W_hh_i+b_h),e_i∈[-1,1] (7)

wherein W_hAs a weight matrix, b_hIs the bias term. Attention weighting parameter alpha_iAnd characteristic quantity h_iMultiplication by multiplication

h_i'＝α_i·h_i (9)

The attention vector H' can be obtained as H ═ H₁',h'₂,...h_t'}。

In the method, a physical method, a regression method and a time series method are the earliest used photovoltaic power generation prediction methods, the three methods have fewer used parameters, the structure is simple, the realization is easy, and the prediction effect is common. The regression method and the time series method mostly use a machine learning algorithm to complete regression and time series prediction at present due to the penetration of a machine learning technology. The BP neural network serving as a typical traditional machine learning algorithm is an unstable model and has the problems of easy falling into local optimization and slow iterative convergence, and the setting of the number of neuron layers of the BP neural network easily causes overfitting and degradation phenomena. The SVM is a stable model and is not easy to fall into local optimization, but the SVM is only suitable for processing the problem of small samples and also has the problem of slow iterative convergence. Traditional machine learning algorithms such as BP neural networks and SVM rely on feature engineering, the prediction accuracy of a model is directly determined by the quality of sample features, the existing artificial design features are still in the domination position, the artificial design features need to depend on the priori knowledge of designers, the model needs to be adjusted manually, only a small number of parameters are allowed to serve as the sample features, photovoltaic power station monitoring data or NWP data are usually and directly used as the feature parameters, the traditional machine learning algorithms cannot effectively extract deeper information in the feature parameters, the prediction model accuracy still has improved space, future photovoltaic power generation amount cannot be predicted relatively accurately, the influence on the stability and safety of a power grid can be caused when large-scale photovoltaic power is merged into the power grid, and photovoltaic power generation prediction is one of key technologies for eliminating and threatening to maintain the safety and stability of the power grid. The reason that prediction accuracy is not high due to traditional machine learning is taken as an example, in the field of huge photovoltaic power generation data quantity, the BP neural network can only construct a shallow neural network, the number of neuron layers is too large, accuracy is affected, characteristics in data are not fully utilized, and the defect that the accuracy is not high is caused compared with deep learning on the whole. Taking an SVM support vector machine as an example, the SVM is suitable for small-scale data analysis, and when the data volume is large, the convergence rate is low, so that the SVM support vector machine is difficult to apply to the field of photovoltaic power generation power prediction.

In summary, the problems of the prior art are as follows: the traditional machine learning algorithm depends on artificially designed sample characteristics, and a short board exists in the aspect of characteristic extraction, so that the accuracy of a prediction model is not high.

The difficulty of solving the technical problems is as follows: at present, the application of the deep learning technology in the photovoltaic industry is still in the initial stage, and documents and technical supports are few. The appearance of deep learning raises the heat of artificial intelligence, and as the latest large field of artificial intelligence at present, the technology of the field is improved by adopting a deep learning algorithm in many other fields. CNN convolutional layers have been used very widely, but the number of filters is not defined exactly by a formula when extracting features. The application of the long and short term memory network LSTM as a typical model structure in the deep learning field to time sequence data analysis is relatively mature, but the photovoltaic data prediction is still in a starting stage, the prior experimental results and experience are less, and the difficulty is increased for overcoming the technical bottleneck. The Attention mechanism has more applications in the fields of image processing, video analysis and natural language processing, but has less applications in the field of photovoltaic power generation prediction.

The significance of solving the technical problems is as follows: the CNN LSTM model combining the CNN, the LSTM and the attention mechanism is a brand-new photovoltaic power generation prediction model, has high feature extraction capability, overcomes the defect of short feature extraction of the traditional machine learning model, improves the prediction precision of photovoltaic power generation power, and further expands the application of the deep learning technology in the field of photovoltaic power generation prediction.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides a photovoltaic power generation power prediction model based on CNN LSTM and a construction method thereof.

The invention is realized in such a way that a photovoltaic power generation power prediction model based on CNN LSTM is composed of a convolution layer CNN, LSTM, an attention mechanism and a full connection layer;

the CNN LSTM model predicts and outputs the photovoltaic power generation power P (t +1) at the t +1 moment according to the photovoltaic power generation power time sequence P (t-size +1), P (t-size +2),.., P (t-1) and P (t), wherein the size is the sequence length.

Another objective of the present invention is to provide a method for constructing the CNN LSTM-based photovoltaic power generation power prediction model, where the method for constructing the CNN LSTM-based photovoltaic power generation power prediction model includes: performing feature extraction on the input photovoltaic power generation time sequence by using 16 convolution kernel one-dimensional convolution layers, and performing feature extraction on the obtained result by using 32 convolution kernel one-dimensional convolution layers; obtaining an LSTM hidden layer output vector by using an LSTM layer with the depth of 32; automatically paying attention to the LSTM hidden layer output vector by using an attention mechanism, and giving a larger weight to the characteristic quantity which is obviously related to the current output quantity; the attention vector is expanded into a one-dimensional vector, the one-dimensional vector is input into a fully-connected layer, and prediction is output by the fully-connected layer.

Further, the construction method of the CNN LSTM-based photovoltaic power generation power prediction model specifically comprises the following steps:

firstly, normalization processing is carried out on photovoltaic power generation power by adopting a normalization method;

secondly, passing the time sequence data through a one-dimensional convolution layer of 16 convolution kernels to obtain a result, and then passing the result through a one-dimensional convolution layer of 32 convolution kernels to obtain a final result, namely the extracted feature;

putting the obtained features into an LSTM layer of each neuron containing 32 LSTMs to obtain LSTM hidden layer output vectors;

step three: automatically paying attention to the LSTM hidden layer output vector by using an attention mechanism, and automatically giving a larger weight to parameters obviously related to the target value;

expanding the feature vector processed by the attention mechanism into a one-dimensional vector so as to meet the requirement of full-connection layer input;

inputting the one-dimensional vector obtained by the unfolding operation into a full-connection layer, and outputting a predicted value of the photovoltaic power generation power at the next moment by the full-connection layer;

step six, taking mean square error MSE as a loss function;

and seventhly, optimizing the weight of the whole neural network through an Rmprop optimizer to minimize the loss value of the loss function.

Further, in the first step, the following normalization method is adopted:

wherein x_iAs the original data, it is the original data,

is normalized data.

Further, in the sixth step, by taking a mean square error MSE as a loss function:

wherein x_modelIs the model output value, x_actualIs an actual measurement value.

The invention also aims to provide a photovoltaic power generation power prediction model based on the CNN LSTM photovoltaic power generation power prediction model, which uses the photovoltaic power generation power time sequence from the current time t to the time t-size +1 collected by a photovoltaic power station monitoring system as the input of the CNN LSTM model to predict the photovoltaic power generation power at the next time.

In summary, the advantages and positive effects of the invention are: as shown in Table 1, the CNN-LSTM used in the experiment has higher prediction precision, and has obvious advantages compared with other traditional machine learning, physical methods and regression methods. Deep learning only requires a simple network structure to realize the approximation of complex functions by learning a deep nonlinear network structure. The deep learning can better extract the characteristics of time series data, and meanwhile, the model has the capability of representing large-scale data due to deep hierarchy and strong expression capability. The CNN convolutional layer is a basic layer for deep learning and common feature extraction. The long-short term memory network is taken as a typical model structure in the field of deep learning, and is widely researched and applied in the aspect of time series analysis, the attention mechanism can automatically give a larger weight to useful parameters, the CNN and the LSTM are combined to automatically extract sample characteristics with high quality, the high-quality characteristics enable the model to have higher prediction precision, and the attention mechanism can effectively improve the generalization capability of the model. The CNN convolution layer can extract various characteristics through a plurality of filters, output vectors of the LSTM hidden layer have hundreds of parameters, and an attention mechanism can automatically endow useful items with larger weights, so that the characteristic extraction quality is further improved, and the high-quality characteristics can improve the photovoltaic power generation power prediction accuracy.

TABLE 1 comparison of the respective model techniques

Note: mean absolute percentage error and root mean square error.

The invention realizes the CNN LSTM model based on Python language and Keras framework, and has the advantages of rapid modeling, high portability and high prediction precision.

The invention is based on a CNN LSTM photovoltaic power generation power prediction model and a construction method and application thereof, the photovoltaic power station historical data is used for training the prediction model, and a verification set is used for evaluating the prediction accuracy of the prediction model in different seasons and different time ranges. The invention realizes a CNN LSTM model based on a convolutional layer (CNN), a long-short term memory network (LSTM) and an Attention mechanism (Attention mechanism) based on a Python language and a Keras framework, wherein the model uses a photovoltaic power generation power time sequence as input and outputs a predicted value of the photovoltaic power generation power at the next moment; CNNs can extract a variety of features of data. The LSTM can realize information abstraction from the input features to high-level features; the attention mechanism can automatically pay attention to the LSTM hidden layer output vector, and a larger weight is given to the characteristic quantity which is obviously related to the current output quantity; and processing the output vector of the attention mechanism into a one-dimensional vector to be input into the full-connection layer through expansion operation, and directly outputting the predicted value of the photovoltaic power generation power at the next moment by the full-connection layer.

The deep learning has flexibility, and the hidden layer can be adjusted according to the actual data volume. The deep learning has excellent prediction capability, and particularly, the accuracy of the power generation amount prediction can be improved for the prediction of large and medium data sets. The invention uses a Keras framework to realize a CNN LSTM model, and applies the CNN LSTM model to the prediction of photovoltaic power generation power.

Drawings

Fig. 1 is a diagram of a CNN LSTM model structure provided in an embodiment of the present invention.

Fig. 2 is a flowchart of a CNN LSTM model construction method provided in the embodiment of the present invention.

Fig. 3 is a schematic diagram of CNN LSTM model training and prediction provided by the embodiment of the present invention.

Fig. 4 is a comparison diagram of the sunny prediction results of the CNN LSTM model provided by the embodiment of the present invention.

Fig. 5 is a diagram comparing the rain prediction results of the CNN LSTM model provided in the embodiment of the present invention.

Fig. 6 is a comparison diagram of the multi-cloud prediction result of the CNN LSTM model provided by the embodiment of the present invention.

FIG. 7 is a comparison graph of the CNN LSTM model and the comparison model MAPE provided by the embodiment of the present invention.

FIG. 8 is a comparison of the CNN LSTM model and the RMSE model provided by the embodiment of the present invention.

FIG. 9 is a comparison between the CNN LSTM model and the MAE model.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

The method aims at the problems that the traditional machine learning algorithm depends on manual design of sample features, and short plates exist in the aspect of feature extraction, so that the precision of a prediction model is not high. The invention realizes the CNN LSTM model based on Python language and Keras framework, and has the advantages of rapid modeling, high portability and high prediction precision.

The following detailed description of the principles of the invention is provided in connection with the accompanying drawings.

As shown in fig. 1, the structure of the CNN LSTM-based photovoltaic power generation power prediction model provided by the embodiment of the present invention mainly includes a convolution layer, an LSTM, an attention mechanism, and a full link layer, and the expansion operation converts an attention vector into a data format that can be processed by the full link layer.

According to the CNN LSTM model provided by the embodiment of the invention, the photovoltaic power generation power time sequence P (t-size +1), P (t-size +2) at the t moment is used as input (the size is the sequence length), P (t-1) and P (t), and the photovoltaic power generation power P (t +1) at the t +1 moment is predicted and output.

As shown in fig. 2, the method for constructing a CNN LSTM-based photovoltaic power generation power prediction model provided in the embodiment of the present invention specifically includes the following steps:

s101: normalization processing is carried out on the photovoltaic power generation power by adopting a normalization method;

s102: the time sequence data passes through a one-dimensional convolution layer of 16 convolution kernels, the obtained result passes through a one-dimensional convolution layer of 32 convolution kernels, and the obtained final result is the extracted feature;

s103: putting the obtained features into an LSTM layer of each neuron containing 32 LSTMs to obtain LSTM hidden layer output vectors;

s104: the feature vectors extracted by the LSTM comprise hundreds of parameters, the LSTM hidden layer output vectors are automatically attended by using an attention mechanism, and the parameters obviously related to the target value are automatically endowed with larger weights, so that the feature extraction quality is further improved;

s105: unfolding the feature vectors processed by the attention mechanism into one-dimensional vectors to meet the requirement of full-connection layer input;

s106: inputting the one-dimensional vector obtained by the unfolding operation into a full-connection layer, and outputting a predicted value of the photovoltaic power generation power at the next moment by the full-connection layer;

s107: taking Mean Square Error (MSE) as a loss function;

s108: and optimizing the weight of the whole neural network through an Rmprop optimizer to minimize the loss value of the loss function.

In a preferred embodiment of the present invention, in step S101, the CNN LSTM-based photovoltaic power generation power prediction model provided in the embodiment of the present invention uses a photovoltaic power generation power time series as an input, the photovoltaic power generation power and the component temperature unit are different, and the daily fluctuation range is large, so that the following normalization methods are adopted:

wherein x_iAs the original data, it is the original data,

is normalized data.

In a preferred embodiment of the present invention, in step S107, the mean square error MSE provided by the embodiment of the present invention is used as a loss function;

wherein x_modelIs the model output value, x_actualIs an actual measurement value.

According to the convolutional layer provided by the embodiment of the invention, the photovoltaic power generation time sequence contains different information, so that the time sequence is put into a one-dimensional convolutional layer with 16 convolutional kernels for feature extraction, the obtained result is put into a 32 convolutional layer for feature extraction, finally obtained features are put into an LSTM to be converted into high-level features, and the output vector of the LSTM hidden layer is the extracted high-level features.

The embodiment of the invention provides an application method of a CNN LSTM-based photovoltaic power generation power prediction model, which specifically comprises the following steps:

the method is characterized in that a CNN LSTM network structure is realized based on Python language and a Keras framework, a photovoltaic power generation power prediction model is established for historical data of a photovoltaic power station, and a photovoltaic power generation power time sequence from the current time t to the time t-size +1, which is acquired by a photovoltaic power station monitoring system, is used as the input of the CNN LSTM model (the size is the length of the time sequence) to predict the photovoltaic power generation power at the next time (the time t + 1).

The application of the principles of the invention will be further illustrated with reference to specific embodiments.

As shown in fig. 3, a training process based on the CNN LSTM photovoltaic power generation power prediction model provided by the embodiment of the present invention is as follows:

(1) data normalization processing: the photovoltaic power generation power has large daily fluctuation range, so normalization processing needs to be respectively carried out, and the method adopts the following normalization method:

wherein x_iAs the original data, it is the original data,

the normalized data is obtained; the training samples at time t may be expressed as:

sample＝[(P(t-size+1),P(t-size+2),...,P(t-1),P(t),P(t+1)]；

wherein P is the photovoltaic power generation power, size is the time sequence length, and P (t +1) is the training target of the training sample;

(2) forward propagation of the CNN LSTM model: inputting a training sample into a CNN LSTM model, performing feature extraction on the CNN and LSTM, distributing weights through an attention mechanism, unfolding an attention vector into a one-dimensional direction, inputting the one-dimensional direction into a full-link layer, and obtaining forward propagation training output;

(3) and (3) calculating a loss value: calculating an MSE loss value according to training output and actual power normalized by the training samples;

wherein x_modelIs the model output value, x_actualIs an actual measurement value;

(4) optimizing the network weight: according to the loss value, optimizing the network weight of the whole CNN LSTM model by using an Rmpp optimizer, and reducing the loss value of the network;

(5) and (4) the steps (2), (3) and (4) are iterated circularly until the loss value is not changed greatly, so that the training of the CNN LSTM model is completed.

The embodiment of the invention provides a prediction process based on a CNN LSTM photovoltaic power generation power prediction model, which comprises the following steps:

(1) data normalization processing: the CNN LSTM model is trained using normalized samples, so that normalization processing needs to be performed on input prediction samples during prediction, and the same normalization method as the training process is adopted.

(2) Forward propagation of the CNN LSTM model: and inputting the normalized prediction sample into the CNN LSTM model to obtain a model output value.

(3) Reverse normalization: the CNN LSTM model is trained by using a normalized sample, so that an output value is also a normalized value, and reverse normalization is required to obtain a true value;

wherein

For the forward propagation output value, x, of the CNN LSTM model_actualThe true value obtained by inverse normalization.

The effect of the present invention will be described in detail with reference to comparative experiments.

The data adopted by the embodiment of the invention is from the measured monitoring data of a certain 20kW photovoltaic power station 2014-2018.

The comparative model adopted in the embodiment of the invention is as follows:

persistence model: the Persistence Model (PM) is a comparison model commonly used for short-term prediction, and considers that the photovoltaic power generation power at the next moment is the same as the power generation power at the current moment, and can be simply expressed as:

wherein

Is the predicted photovoltaic power generation function at the next momentThe rate, p (t), is the photovoltaic power generation power at the present moment.

ARIMA model: the ARIMA is a common time series prediction model in statistical models and is called an autoregressive integral moving average model.

MLP model: the MLP model takes a generating power time sequence and a component temperature time sequence as input, and outputs a photovoltaic generating power predicted value at the next moment through two hidden layers.

LSTM model: the LSTM model respectively extracts the characteristics of the input power generation time sequence and the component temperature time sequence by using two LSTM layers, expands the extracted characteristic vectors into one dimension, then fuses the vectors, and outputs a prediction result through a full connection layer.

The CNN LSTM model (CLSTM) of the embodiment of the invention performs 7.5min, 15min, 30min and 60min prediction results in spring, summer, autumn and winter as shown in FIGS. 3-6. The Persistence Model (PM) and the ARIMA model have advantages in prediction in 15min and below, but when prediction is carried out for 30min and 60min, prediction curves of the PM and the ARIMA obviously deviate from measured values. The MLP model and the LSTM model can make effective prediction in 30min and below, and a certain difference still exists between the MLP model and the LSTM model and compared with the CNN LSTM model. The CNN LSTM model can effectively predict the photovoltaic power generation power in spring, summer, autumn and winter, the prediction of the model in 7.5min, 15min and 30min is very close to the measured value, the CNN LSTM model can still grasp the trend of the photovoltaic power generation power change in the prediction of 60min, and the prediction curve is closer to the actual measurement curve than the four comparison models.

The embodiment of the invention adopts the average absolute percentage error MAPE, the root mean square error RMSE and the average absolute error MAE to evaluate the prediction performance of the model.

Wherein x_modelAs model predicted values, x_actualIs an actual measurement value.

The MAPE ratio of the CNN LSTM model (CLSTM) of the present example and the four comparative models for annual prediction is shown in table 2.

TABLE 2 annual MAPE comparison

		2017.10	2017.11	2017.12	2018.1	2018.2	2018.3	2018.4	2018.5	2018.6	2018.7	2018.8	2018.9	Average
																PM	23.22	28.11	17.9	23.92	21.96	22.15	18.43	26.2	22.74	20.77	22.02	25.88	22.78
	MLP	32.25	32.05	25.83	40.5	35.39	31.13	28.22	35.88	35.39	31.71	30.45	32.62	32.62
															7.5min	LSTM	31.39	33.41	27.88	45.02	31.71	31.28	19.92	29.6	30.84	30	26.25	31.28	30.71
	CNN	31.41	33.45	24.7	38.9	32.55	30.43	20.93	33.57	31.92	26.73	24.68	29.36	29.89
																The invention	28.63	33.94	21.29	32.58	29.68	28.11	18.91	32.99	28.27	25.68	27.98	29.03	28.09
																PM	28.48	32.69	25.49	30.65	27.57	28.76	25.14	32.79	29.25	26.84	29.68	32.19	29.13
	MLP	28.93	36.97	31.28	45.83	40.25	36.82	30.48	40.72	35.76	37.55	32.67	34.71	36
															15min	LSTM	34.32	35.54	27.23	48.93	33.7	34.83	25.8	38.74	33.04	31.34	30.94	31.56	33.83
	CNN	32.41	34.25	23.66	39.05	32.74	31.72	24.32	35.81	34.51	28.26	27.85	31.33	31.33
																The invention	28.95	35.08	25.38	40.61	33.32	30.83	23.2	35.1	31.64	28.22	27.16	32.5	31
																PM	36.84	39.53	34.18	38.91	38.21	36.43	31.8	41.09	37.01	33.13	37.37	42.7	37.27
	MLP	41.94	40.57	33.99	48.41	44.27	45	36.24	41.65	41.17	39.03	42.84	41.58	41.39
															30min	LSTM	37.96	36.32	33.13	45.82	38.93	37.86	29.62	43.03	38.34	36.13	38.86	39.21	37.94
	CNN	34.63	37.22	26.75	45.41	37.81	32.54	27.78	38.42	34.72	32.5	30.27	35.76	34.48
																The invention	29.28	41.22	25.98	45.37	38.84	33.61	26.95	38.68	33.49	30.75	25.75	37.28	33.93
																PM	50.7	50.5	46.94	50.12	46.68	49.93	43.16	48.61	45.31	43.46	46.93	48	47.53
	MLP	44.89	47.03	39.7	56.47	51.23	47.46	41.51	50.36	47.32	43.23	45.55	44.48	46.6
															60	LSTM	44.33	46.59	28.76	53.04	40.08	36.99	38.03	49.22	39.87	40.34	44.27	39.32	41.74
	CNN	35.92	43.1	36.93	46.4	31.36	37.7	37.79	37.7	34.52	37.5	38.51	41.79	38.27
																The invention	33.65	29.51	25.36	49.88	38.11	37.63	30.35	42.48	37.8	32.86	26.41	30.85	34.57

The RMSE pair ratios of the CNN LSTM model (CLSTM) of the present invention embodiment and the four comparative models for annual prediction are shown in table 3.

TABLE 3 annual RMSE comparison

The MAE alignment of the CNN LSTM model (CLSTM) of the present invention and the four comparative models for annual prediction is shown in Table 4.

TABLE 4 annual MAE comparison

		2017. 10	2017. 11	2017. 12	2018. 1	2018. 2	2018. 3	2018. 4	2018. 5	2018. 6	2018. 7	2018. 8	2018. 9	Average
																PM	0.81	0.96	0.44	0.47	0.58	0.69	0.79	0.97	0.81	0.86	1.08	1.32	0.82
	MLP	1.08	0.98	0.75	0.59	0.99	0.9	1.54	1.19	1.35	1.29	1.42	1.46	1.13
															7.5	LSTM	0.8	0.93	0.6	0.54	0.57	0.73	0.77	0.98	0.84	1.08	1.12	1.27	0.85
	CNN	0.82	0.94	0.62	0.45	0.53	0.65	0.72	0.98	0.82	0.82	1.07	1.25	0.8
																Hair brush Ming dynasty	0.73	0.91	0.39	0.45	0.51	0.67	0.71	0.99	0.79	0.85	1.23	1.25	0.79
																PM	1.03	1.06	0.68	0.67	0.81	0.93	1.02	1.26	1.08	1.12	1.43	1.67	1.06
	MLP	1.24	1.05	0.6	0.72	0.8	1.02	1.5	1.31	1.26	1.95	1.48	1.54	1.21
															15	LSTM	0.93	1.03	0.51	0.62	0.67	0.86	0.96	1.2	1.03	1.2	1.36	1.56	0.99
	CNN	0.9	1	0.47	0.55	0.66	0.79	1.03	1.2	1.17	1.01	1.26	1.45	0.96
																Hair brush Ming dynasty	0.86	1.03	0.46	0.57	0.64	0.79	0.86	1.23	1.01	1.04	1.24	1.48	0.93
																PM	1.36	1.29	1.07	0.96	1.22	1.33	1.54	1.68	1.5	1.56	1.94	2.35	1.48
	MLP	1.16	1.13	0.83	0.79	1.02	1.55	1.64	1.52	1.34	1.86	2.26	2.13	1.43
															30	LSTM	1.26	1.08	0.77	0.78	0.93	1.05	1.23	1.47	1.37	1.49	1.81	1.95	1.27
	CNN	1.1	1.11	0.64	0.71	0.94	1.04	1.11	1.53	1.26	1.4	1.46	2.07	1.2
																Hair brush Ming dynasty	1.22	1.16	0.72	0.71	1.04	0.93	1.04	1.45	1.24	1.29	1.45	1.91	1.18
																PM	2.04	1.68	1.83	1.47	1.83	2.09	2.38	2.19	1.97	2.35	2.62	3.02	2.12
	MLP	1.46	1.32	1.08	1.04	1.48	1.58	2.06	1.8	1.73	2.07	2.23	2.67	1.71
															60	LSTM	1.43	1.4	0.94	1.04	1.34	1.41	1.83	1.87	1.59	2.05	2.27	2.43	1.63
	CNN	1.45	1.32	1.18	1.06	1.18	1.3	1.84	1.69	1.57	1.92	1.9	2.48	1.57
																Hair brush Ming dynasty	1.26	1.26	0.94	1.03	1.31	1.7	1.3	1.6	1.29	1.42	1.32	1.93	1.36

From the comparative analysis of the experimental data in tables 3 and 4, it can be seen that: the predicted mean MAPE of the Persistence Model (PM) was lowest at 7.5min, and the predicted mean MAPE of the Persistence model was greatest at one hour as the time horizon expanded; the ARIMA model has the advantages that the average MAPE is better than that of the MLP model and the LSTM model when predicted at 7.5min, and the average MAPE, RMSE and MAE predicted at other periods are higher than that of the MLP model, the LSTM model and the CNN LSTM model (CLSTM); when the MLP model is predicted for 7.5min, the average MAPE of the MLP model is very similar to that of the LSTM model, the disadvantages of the MLP model are gradually revealed along with the increase of the prediction time range, and the average values of three evaluation indexes of the MLP model are always larger than that of the LSTM model during prediction for 15min, 30min and 60min because the MLP model does not have the capability of time series feature extraction; the LSTM model has strong time sequence feature extraction capability, has obvious advantages when being predicted in a large time range compared with a traditional ARIMA time sequence prediction model, and also has certain advantages on three evaluation indexes compared with an MLP model; the CNN LSTM model is an improved model added with an attention mechanism on the basis of the LSTM model, compared with the LSTM model, the MAPE of the CNN LSTM model is obviously reduced, the RMSE and the MAE are also reduced slightly, and the CNN LSTM model has better effects in prediction of 7.5min, 15min, 30min and 60 min.

The broken lines of MAPE, RMSE and MAE predicted by the CNN LSTM model (CLSTM) and the four comparison models in different time ranges of the embodiment of the invention are shown in FIGS. 7-9.

As shown in FIG. 7, a comparison graph of the CNN LSTM model and the MAPE model is provided.

As shown in FIG. 8, a comparison graph of the CNN LSTM model and the RMSE model is provided.

As shown in fig. 9, a comparison graph of the CNN LSTM model and the comparison model MAE is provided in the embodiment of the present invention.

From the MAPE broken line comparison of FIG. 7, it can be seen that the Persistence Model (PM) predicts the lowest MAPE at 7.5min, the MAPE of the Persistence model gradually increases with the expansion of the prediction time range, the prediction result of the Persistence model at 60min has little reference value, and the MAPE curve of the CNN LSTM model (CLSTM) is generally lower than that of the other four reference models. From the comparison of the broken lines in FIGS. 8 and 9, it can be seen that the predicted RMSE and MAE of the CNN LSTM model at 7.5min, 15min, 30min and 60min are significantly lower than the four comparative models.

In a comprehensive view, the CNN LSTM model obtains better prediction performance in cloudy days in sunny days and rainy days and in four prediction time ranges of 7.5min, 15min, 30min and 60 min.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.

Claims (3)

1. A construction method of a photovoltaic power generation power prediction model based on CNN LSTM is characterized in that the photovoltaic power generation power prediction model based on CNN LSTM is composed of a convolution layer CNN, LSTM, an attention mechanism and a full connection layer;

the CNN LSTM model predicts and outputs the photovoltaic power generation power P (t +1) at the t +1 moment by taking the photovoltaic power generation power time sequence P (t-size +1), P (t-size +2),. the P (t-1) and P (t), wherein the size is the sequence length;

the construction method of the photovoltaic power generation power prediction model based on the CNN LSTM comprises the following steps: performing feature extraction on the input photovoltaic power generation time sequence by using 16 convolution kernel one-dimensional convolution layers, and performing feature extraction on the obtained result by using 32 convolution kernel one-dimensional convolution layers; obtaining an LSTM hidden layer output vector by using an LSTM layer with the depth of 32; automatically paying attention to the LSTM hidden layer output vector by using an attention mechanism, and giving a larger weight to the characteristic quantity which is obviously related to the current output quantity; unfolding the attention vector into a one-dimensional vector, inputting the one-dimensional vector into a full-link layer, and outputting prediction by the full-link layer;

the construction method of the CNN LSTM-based photovoltaic power generation power prediction model specifically comprises the following steps:

step one, carrying out normalization processing on photovoltaic power generation data;

secondly, passing the time sequence data through a one-dimensional convolution layer of 16 convolution kernels to obtain a result, and then passing the result through a one-dimensional convolution layer of 32 convolution kernels to obtain a final result, namely the extracted feature;

putting each obtained dimensional feature into an LSTM neuron containing 32 LSTMs and consisting of multidimensional features to obtain an LSTM hidden layer output vector;

step three: automatically paying attention to the LSTM hidden layer output vector by using an attention mechanism, and automatically giving a larger weight to parameters obviously related to the target value;

expanding the feature vector processed by the attention mechanism into a one-dimensional vector so as to meet the requirement of full-connection layer input;

inputting the one-dimensional vector obtained by the unfolding operation into a full-connection layer, and outputting a predicted value of the photovoltaic power generation power at the next moment by the full-connection layer;

step six, taking mean square error MSE as a loss function;

and seventhly, optimizing the weight of the whole neural network through an Rmprop optimizer to minimize the loss value of the loss function.

2. The method for constructing a CNN LSTM-based photovoltaic power generation power prediction model according to claim 1, wherein in the first step, the following normalization method is adopted:

wherein x_iAs the original data, it is the original data,

is normalized data.

3. The method for constructing a CNN LSTM-based photovoltaic power generation power prediction model according to claim 1, wherein in the sixth step, by using a mean square error MSE as a loss function:

wherein x_modelIs the model output value, x_actualIs an actual measurement value.

Images