A photovoltaic power station power generation day-ahead prediction method based on spatial downscaling and TCN-BiLSTM

Abstract

The application discloses a photovoltaic power station power generation power day-ahead prediction method based on spatial scale reduction and TCN-BiLSTM, and belongs to the field of new energy power generation power prediction. Historical power, historical ground meteorological observation and numerical weather prediction data of a target photovoltaic station and reference stations are acquired, and temperature, global horizontal irradiance and direct irradiance spatial deviation are calculated; historical power before a prediction release time, historical spatial deviation and prediction day numerical weather prediction data are combined to form a multi-source input, and a TCN-BiLSTM model optimized by Bayesian optimization is used to output a prediction day power sequence. Historical measured data is only taken from before the prediction release time, and future measured meteorological data of the prediction day is not used. The method can supplement local deviation information of the target station relative to a regional meteorological background, and reduce day-ahead prediction errors under complex weather conditions.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation and power system operation control technology, specifically to a day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM. Background Technology

[0002] As the installed capacity and grid connection rate of photovoltaic power generation increase, the volatility and uncertainty of photovoltaic power plant output power increase the difficulty of day-ahead dispatching, reserve capacity allocation, and power balance control of the power system. Day-ahead power forecasting of photovoltaic power plants typically uses historical power generation data, surface meteorological observation data, and numerical weather prediction data to estimate the power generation curve for the forecast day.

[0003] Some existing methods primarily extrapolate time series data based on the historical power output of the target photovoltaic (PV) power plant. When rapid cloud movement, localized rainfall, or sudden changes in irradiance occur during the forecast period, historical power output may not adequately reflect future meteorological disturbances, easily leading to lag in the forecast curve, peak bias, and omission of local fluctuations. Other methods use numerical weather prediction (NMR) data as model input; however, NMR typically describes the macro-meteorological state of a region using a large spatial grid, making it difficult to fully reflect local temperature and irradiance changes near specific PV power plants.

[0004] Multiple photovoltaic (PV) power plants within the same region are affected by similar large-scale weather systems. However, due to differences in geographical location, cloud distribution, and local environment, the meteorological conditions of the target PV power plant will dynamically deviate from the regional average meteorological background. Existing multi-station prediction methods often employ data splicing from neighboring stations, overall aggregation, or spatial correlation learning, which may not be able to represent the degree of deviation of the target power plant from the regional meteorological background with features that have clear physical meaning. Furthermore, PV power changes under complex weather conditions exhibit both short-term abrupt changes and long-term evolutionary trends, making it difficult for a single convolutional network or unidirectional temporal series network to simultaneously extract these temporal features.

[0005] Therefore, there is a need for a day-ahead power prediction method that can construct local spatial bias characteristics using historical observation data from multiple stations and jointly model the historical local conditions before the prediction release time with the numerical weather forecast of the prediction day, in order to reduce prediction errors under complex weather conditions. Summary of the Invention

[0006] The purpose of this invention is to provide a day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM, executed by a power prediction device, comprising: Historical sample data is obtained, including historical power generation data of the target photovoltaic power station, historical surface meteorological observation data of the target photovoltaic power station and multiple reference photovoltaic power stations in its area, and numerical weather forecast data corresponding to the historical day-ahead period; The regional meteorological background is determined based on the historical surface meteorological observation data of the target photovoltaic power station and the multiple reference photovoltaic power stations. The spatial deviation of temperature, the spatial deviation of global horizontal irradiance, and the spatial deviation of direct irradiance are determined based on the differences between the historical surface meteorological observation values ​​of the target photovoltaic power station and the corresponding regional meteorological background. The historical power generation sequence and the spatial deviation sequence located within the historical input period are combined with the numerical weather forecast sequence of the corresponding historical day-ahead period to form historical training samples. The actual power generation sequence of the historical day-ahead period is used as the label to train a combined prediction model including a temporal convolutional network, a bidirectional long short-term memory network and a fully connected output layer. When executing the forecast, obtain the historical power generation sequence and historical spatial deviation sequence of the target photovoltaic power station before the forecast release time, and obtain the numerical weather forecast sequence corresponding to the forecast day; According to the same feature order as the historical training samples, the historical power generation sequence, the historical spatial deviation sequence, and the numerical weather forecast sequence corresponding to the prediction day are combined into a multi-source input sample to be predicted; The combined prediction model, trained by inputting the multi-source input samples to be predicted, extracts local temporal variation features through the temporal convolutional network, extracts bidirectional temporal dependency features through the bidirectional long short-term memory network, and outputs the day-ahead power generation prediction sequence of the target photovoltaic power station through the fully connected output layer.

[0008] Furthermore, the historical power generation data, the historical surface meteorological observation data, and the numerical weather forecast data are preprocessed, including: removing data from periods without sunshine and retaining valid sunshine period data from 6:00 to 21:00 each day; matching the data according to valid timestamps; and normalizing the data in different dimensions using the maximum-minimum normalization method.

[0009] Furthermore, the temporal clarity index is calculated based on the predicted total horizontal irradiance from historical numerical weather forecast data and the historical measured global horizontal irradiance at the corresponding time. A first preset weather classification threshold and a second preset weather classification threshold are set, with the first preset weather classification threshold being greater than the second preset weather classification threshold. When the temporal clarity index is greater than or equal to the first preset weather classification threshold, the corresponding historical time is classified as a clear weather condition. When the temporal clarity index is greater than or equal to the second preset weather classification threshold and less than the first preset weather classification threshold, the corresponding historical time is classified as a cloudy weather condition. When the temporal clarity index is less than the second preset weather classification threshold, the corresponding historical time is classified as a rainy weather condition.

[0010] Furthermore, each historical moment and test moment is assigned to a corresponding weather condition subset according to the aforementioned weather conditions. The weather condition subsets are used to evaluate the prediction performance of the combined prediction model under sunny, cloudy, and rainy conditions, respectively.

[0011] Furthermore, the regional meteorological background includes the average temperature, average global horizontal irradiance, and average direct irradiance of the target photovoltaic power station and the multiple reference photovoltaic power stations at corresponding historical times; the spatial deviation of temperature is the difference between the historical measured temperature of the target photovoltaic power station and the average temperature, the spatial deviation of global horizontal irradiance is the difference between the historical measured global horizontal irradiance of the target photovoltaic power station and the average global horizontal irradiance, and the spatial deviation of direct irradiance is the difference between the historical measured direct irradiance of the target photovoltaic power station and the average direct irradiance.

[0012] Furthermore, the historical input period is located before the corresponding historical day-ahead period; the historical input period and the historical day-ahead period have the same sampling interval and the same number of sampling points, and the historical power generation sequence, the spatial deviation sequence and the numerical weather forecast sequence are combined according to the same sampling sequence number; when performing day-ahead forecasting, the time endpoint of the historical power generation sequence and the historical spatial deviation sequence is not later than the forecast release time.

[0013] Furthermore, the bidirectional long short-term memory network includes forward long short-term memory units and reverse long short-term memory units. The forward long short-term memory units process the features output by the temporal convolutional network in forward time order of the input sequence, and the reverse long short-term memory units process the features output by the temporal convolutional network in reverse time order of the input sequence, and concatenate the forward hidden state and the reverse hidden state at the same time step; the fully connected output layer outputs the day-ahead normalized power generation prediction value based on the concatenated bidirectional hidden state.

[0014] Furthermore, a Bayesian optimization algorithm is used to optimize the hyperparameters of the combined prediction model. The hyperparameters include the number of network channels of the temporal convolutional network, the number of hidden layer neurons of the bidirectional long short-term memory network, the initial learning rate, the residual block dropout rate, and the batch sample processing number.

[0015] Furthermore, the root mean square error is used as the loss function of the combined prediction model, and an adaptive moment estimation optimizer is used to update the network parameters of the combined prediction model.

[0016] Furthermore, the temporal convolutional network has 32 network channels, the bidirectional long short-term memory network has 64 hidden layer neurons, the initial learning rate is 0.0024, the residual block dropout rate is 0.3, and the batch sample processing number is 128.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention utilizes historical ground meteorological observation data from the target photovoltaic power station and a reference photovoltaic power station to construct regional meteorological background and three types of spatial deviation characteristics, enabling the model to characterize the historical local meteorological deviation of the target power station relative to the regional average state; 2. In the day-ahead forecasting phase, this invention only uses historical power and historical spatial bias that have been formed before the forecast release time, as well as numerical weather forecast data for the forecast day, without using future ground-based measured meteorological data for the forecast day, thus ensuring that the availability of data for training and online forecasting remains consistent. 3. This invention utilizes a temporal convolutional network to extract local temporal variation features and a bidirectional long short-term memory network to extract bidirectional temporal dependencies of the same input sequence, enabling the model to jointly characterize short-term meteorological disturbances and power evolution over longer time scales; 4. This invention utilizes the Bayesian optimization algorithm to optimize the hyperparameters of the combined prediction model, reducing the impact of manual parameter configuration on the model's prediction performance; 5. In the embodiments described in this specification, the root mean square error of the TCN-BiLSTM model using only historical power is 0.0383; after adding numerical weather prediction, it is reduced to 0.0142; after further adding spatial bias features, it is reduced to 0.0113, and the mean absolute error is reduced to 0.0085. Attached Figure Description

[0018] Figure 1 A flowchart of a day-ahead prediction method for photovoltaic power generation provided in an embodiment of the present invention.

[0019] Figure 2 This is a geographical distribution map of various photovoltaic power stations provided for an embodiment of the present invention.

[0020] Figure 3This is a typical daily power generation curve of various photovoltaic power plants provided in an embodiment of the present invention.

[0021] Figure 4 This is a comparison chart of day-ahead power generation prediction results when only historical power generation data is used, provided as an embodiment of the present invention.

[0022] Figure 5 A thermogram showing the correlation coefficient between heterogeneous meteorological characteristics and photovoltaic power generation is provided as an embodiment of the present invention.

[0023] Figure 6 This is a comparison chart of day-ahead power generation prediction results before and after incorporating numerical weather forecast data, provided as an embodiment of the present invention.

[0024] Figure 7 This is a comparison chart of day-ahead power generation prediction results before and after incorporating spatial deviation features, provided as an embodiment of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the implementation process of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or conventional adjustments made by those skilled in the art to the following embodiments without departing from the technical concept of the present invention should all fall within the scope of protection of the present invention.

[0026] This invention provides a day-ahead forecasting method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM. The spatial downscaling method refers to using the historical average meteorological conditions of multiple photovoltaic power plants within a region as the regional meteorological background, constructing spatial deviations of temperature, global horizontal irradiance, and direct irradiance of the target photovoltaic power plant relative to the regional meteorological background, and then using these spatial deviations as local micrometeorological characteristics input into the power prediction model.

[0027] In the current forecast phase, the spatial deviation is calculated using historical surface meteorological observation data obtained before the forecast release time. Future meteorological information for the forecast period is provided by numerical weather prediction data, without using actual surface meteorological data that has not yet been generated for the forecast period.

[0028] Example 1 This embodiment provides a day-ahead forecasting method for photovoltaic power plant output. (See also...) Figure 1 The method includes multi-source data acquisition and preprocessing, weather condition classification, spatial deviation feature construction, training sample construction, TCN-BiLSTM combined prediction model training, model hyperparameter optimization, and day-ahead power prediction.

[0029] This embodiment is implemented using the Python programming language and the PyTorch deep learning framework, and validated using the PVOD public dataset. Historical power generation data, surface meteorological observation data, and numerical weather prediction data for approximately one year from five photovoltaic power plants were selected, with a time resolution of 15 minutes. Among them, plant 3 has an installed capacity of 15 MW, and the other four plants each have an installed capacity of 20 MW. The geographical distribution of each photovoltaic power plant is shown in [reference needed]. Figure 2 For typical daily power generation curves of each photovoltaic power station, please refer to [link / reference]. Figure 3 .

[0030] I. Multi-source data acquisition and preprocessing One photovoltaic power station is selected from the five photovoltaic power stations as the target photovoltaic power station, and the remaining photovoltaic power stations are used as reference photovoltaic power stations located around the target photovoltaic power station.

[0031] The system acquires historical power generation data of the target photovoltaic power station, historical surface meteorological observation data of the target photovoltaic power station and the reference photovoltaic power station, and numerical weather forecast data corresponding to the historical surface meteorological observation data in time.

[0032] The surface meteorological observation data includes at least temperature, global horizontal irradiance, and direct irradiance. The numerical weather prediction data is used to characterize the macro-meteorological conditions during the forecast period.

[0033] Based on the physical characteristic that photovoltaic systems do not generate power at night when there is no sunlight, redundant data during periods of no sunlight are removed, and effective sunlight data from 6:00 to 21:00 each day is retained.

[0034] Historical power generation data, historical surface meteorological observation data, and numerical weather prediction data are correlated according to timestamps to give data from different sources a unified time resolution.

[0035] The max-min normalization method is used to normalize data of different dimensions, and its calculation formula is as follows:

[0036] In the formula, For the first in the dataset One set of raw data, and These represent the maximum and minimum values ​​of the corresponding variables. This is the normalized data.

[0037] During model training, the maximum and minimum values ​​of each variable are saved. After obtaining the normalized power prediction result, a mapping method that is the reverse of the normalization process described above is used to restore the normalized power prediction value to the power prediction value under the original power dimension.

[0038] II. Weather Condition Classification To analyze the prediction performance of the combined prediction model under different weather conditions, a weather state discrimination function was constructed based on the irradiance prediction values ​​in numerical weather prediction data and the corresponding historical measured ground irradiance.

[0039] First, calculate the temporal clarity index:

[0040] In the formula, For time-series clarity index, This refers to the predicted horizontal total irradiance output by the numerical weather prediction system. This represents the measured global horizontal irradiance on the ground at the corresponding time. This is used to avoid computational singularities when the measured global horizontal irradiance on the ground is zero.

[0041] Set the first preset weather classification threshold. Second preset weather classification threshold Among them, the first preset weather classification threshold Greater than the second preset weather classification threshold Based on the time-series clarity index and the first preset weather classification threshold... and the second preset weather classification threshold The relationships between these conditions divide historical moments into sunny, cloudy, and rainy weather conditions:

[0042] In the formula, For a moment Corresponding weather conditions For a moment The corresponding temporal clarity index, The first preset weather classification threshold, The second preset weather classification threshold, and .

[0043] In this embodiment, the weather condition classification results are used to perform weather condition-specific statistics and performance evaluation on historical samples and test results, and are not used as input features that must be calculated based on the future measured irradiance during the forecast period in the day-ahead forecast stage.

[0044] III. Construction of Spatial Deviation Features To address the issue that numerical weather prediction has a large spatial coverage area and cannot directly reflect the local micro-meteorological changes of the target photovoltaic power station, the average meteorological state of the target photovoltaic power station and the reference photovoltaic power station at the same historical moment is used as the regional meteorological background to construct the spatial deviation characteristics of the target photovoltaic power station relative to the regional meteorological background.

[0045] The spatial deviation characteristics include temperature spatial deviation, global horizontal irradiance spatial deviation, and direct irradiance spatial deviation.

[0046] Temperature space deviation is calculated using the following formula:

[0047] In the formula, For temperature space deviation, The measured temperature of the target photovoltaic power station at the corresponding historical moment. The average measured temperature of the target photovoltaic power station and the reference photovoltaic power station at this historical moment.

[0048] The spatial deviation of global horizontal irradiance is calculated according to the following formula:

[0049] In the formula, For the spatial bias of global horizontal irradiance, The measured global horizontal irradiance of the target photovoltaic power station at the corresponding historical moment. The average measured global horizontal irradiance of the target photovoltaic power station and the reference photovoltaic power station at this historical moment.

[0050] The spatial deviation of direct irradiance is calculated according to the following formula:

[0051] In the formula, For the spatial bias of direct irradiance, The measured direct irradiance of the target photovoltaic power station at the corresponding historical moment is given. This represents the average measured direct irradiance of the target photovoltaic power station and the reference photovoltaic power station at this historical moment.

[0052] The above processing distinguishes between overall regional meteorological changes and local meteorological deviations of the target photovoltaic power station. When local cloud cover, local rainfall, or local temperature changes occur near the target photovoltaic power station, the three types of spatial deviations can indicate the degree of deviation of the target photovoltaic power station from the regional average meteorological conditions.

[0053] IV. Construction of Training Samples and Current Prediction Samples Training samples for a combined prediction model are constructed using historical date prediction tasks.

[0054] Select historical date tag time period And select the time period tagged before the historical date. Previous historical input periods The historical input period and historical date tags The same 15-minute sampling interval and the same number of sampling points were used.

[0055] For the Each sampling sequence number represents a historical input time period. The historical power generation, temperature spatial deviation, global horizontal irradiance spatial deviation, and direct irradiance spatial deviation of the target photovoltaic power station corresponding to the sampling sequence number are compared with the historical date-tagged time period. The numerical weather forecast meteorological features corresponding to the sampling sequence number are combined to form a multi-source input feature vector:

[0056] In the formula, Input historical time periods Inner The target photovoltaic power generation power of the photovoltaic power station at each sampling time; , and These are the historical input time periods. Inner Spatial deviation of temperature at each sampling time, spatial deviation of global horizontal irradiance, and spatial deviation of direct irradiance; Label the period before the historical date Inner The meteorological feature vector of numerical weather forecast at each sampling time.

[0057] Label the historical date with the time period The actual power generation of the target photovoltaic power station at each sampling time is used as the training label corresponding to the multi-source input feature vector.

[0058] The multi-source input feature vectors are arranged sequentially according to the sampling sequence number to form a multivariate time series input of a historical training sample.

[0059] When performing actual day-ahead forecasts, historical input time period data obtained before the forecast release time are selected, and the spatial deviation of temperature, spatial deviation of global horizontal irradiance, and spatial deviation of direct irradiance within the historical input time period are calculated. Numerical weather forecast meteorological characteristics of each sampling time within the day-ahead period to be forecast are obtained.

[0060] The multi-source input sequence to be predicted is constructed using the same sampling sequence and feature arrangement as in the model training phase. The day-ahead forecasting phase does not use actual power generation data or ground-based meteorological data that have not yet been generated within the day-ahead period to be predicted.

[0061] V. Construction of the TCN-BiLSTM Combined Prediction Model The combined prediction model consists of a temporal convolutional network layer, a bidirectional long short-term memory network layer, and a fully connected output layer.

[0062] A multi-source input sequence consisting of historical power generation, three types of spatial bias, and numerical weather prediction meteorological characteristics is input into a temporal convolutional network.

[0063] The temporal convolutional network is used to extract local temporal variation features between consecutive sampling times from a multi-source input sequence. In this embodiment, the number of network channels of the temporal convolutional network is set to 32, and the dropout rate of the residual blocks of the temporal convolutional network is set to 0.3.

[0064] The features output by the temporal convolutional network are input into a bidirectional long short-term memory (LSTM) network. The bidirectional LSTM network includes forward LSTM units and backward LSTM units.

[0065] The forward long short-term memory unit processes the features output by the temporal convolutional network in forward time order along the input sequence, while the reverse long short-term memory unit processes the features output by the temporal convolutional network in reverse time order along the same input sequence, and then concatenates the forward and reverse hidden states:

[0066] In the formula, For temporal convolutional networks at time steps Output characteristics For positive long short-term memory units at time steps The hidden state of the output. For reverse long short-term memory units at time steps The hidden state of the output. This is a bidirectional temporal feature obtained by concatenating the forward hidden state and the reverse hidden state.

[0067] The reverse long short-term memory unit processes the complete multi-source input sequence that has been input into the combined prediction model, without using the actual power generation of the day before the forecast or future ground-measured meteorological data.

[0068] In this embodiment, the number of hidden layer neurons in the bidirectional long short-term memory network is set to 64.

[0069] The fully connected output layer receives the bidirectional temporal features from the bidirectional long short-term memory network output and obtains the normalized power prediction value through linear weighting and bias mapping:

[0070] In the formula, For time steps The corresponding normalized power prediction value, The weights of the fully connected output layer. This represents the bidirectional temporal characteristics output by a bidirectional long short-term memory network. This is the bias for the fully connected output layer.

[0071] The normalized power prediction values ​​corresponding to each time step are arranged in chronological order to form the day-ahead normalized power generation prediction sequence of the target photovoltaic power station, and the day-ahead power generation prediction curve under the original power dimension is obtained by inverse normalization processing.

[0072] VI. Model Hyperparameter Optimization and Training A Bayesian optimization algorithm is used to automatically optimize the hyperparameters of the combined prediction model based on a probabilistic substitution model.

[0073] The hyperparameters to be optimized include: The number of network channels in a temporal convolutional network, the number of hidden layer neurons in a bidirectional long short-term memory network, the initial learning rate of the optimizer, the dropout rate of residual blocks in a temporal convolutional network, and the number of batch samples processed during model training.

[0074] Within the preset hyperparameter search boundary, candidate hyperparameter combinations are determined. Each candidate hyperparameter combination is used to train the combined prediction model and calculate the corresponding prediction error. Based on the evaluated candidate hyperparameter combinations and prediction errors, subsequent candidate hyperparameter combinations are determined, and the hyperparameter combination with the smaller prediction error is selected from the candidate hyperparameter combinations.

[0075] The combined prediction model uses root mean square error as the loss function:

[0076] In the formula, The number of samples used in a single loss calculation. For the first The per-unit value of the actual power generation of each sample For the first The per-unit value of the predicted power generation for each sample.

[0077] An adaptive moment estimator optimizer is used to update the network parameters of the combined prediction model.

[0078] The first moment estimate of the exponential moving average based on historical gradients is updated according to the following formula:

[0079] The second moment estimate based on the exponential moving average of historical gradient squares is updated according to the following formula:

[0080] Combined with the bias correction mechanism, the network parameters are updated according to the following formula:

[0081] In the formula, The gradient calculated for the current training round. and These are the exponential decay rates for the first-moment estimate and the second-moment estimate, respectively. and These are the first-order moment estimates and second-order moment estimates after bias correction, respectively. The current learning rate, To prevent positive numbers with a denominator of zero, and These are the network parameters before and after the update, respectively.

[0082] The hyperparameter combination used in this embodiment, obtained through Bayesian optimization algorithm, is as follows: the number of network channels of the temporal convolutional network is 32, the number of hidden layer neurons of the bidirectional long short-term memory network is 64, the initial learning rate is 0.0024, the residual block dropout rate is 0.3, and the batch sample processing number is 128.

[0083] VII. Day-ahead Power Generation Forecast At the time of prediction release, historical power generation data of the target photovoltaic power station before the time of prediction release are obtained, historical ground meteorological observation data of the target photovoltaic power station and the reference photovoltaic power station before the time of prediction release are obtained, and the three types of spatial deviations of the corresponding historical input period are calculated according to formulas (4) to (6).

[0084] Obtain the meteorological characteristics of numerical weather forecasts for the period before the forecast date.

[0085] According to the data organization relationship shown in formula (6-1), the historical power generation, three types of historical spatial bias and the numerical weather forecast meteorological characteristics of the day-ahead period to be predicted are combined into the multi-source input sequence to be predicted.

[0086] The multi-source input sequence to be predicted is input into the trained TCN-BiLSTM combined prediction model, and then passed through a temporal convolutional network layer, a bidirectional long short-term memory network layer, and a fully connected output layer to obtain the normalized power prediction value corresponding to each prediction time.

[0087] The normalized power prediction values ​​are inversely normalized and arranged in chronological order of prediction time to obtain the day-ahead power generation prediction curve of the target photovoltaic power station.

[0088] An optional embodiment, based on embodiment 1, uses different input feature configurations and different combined prediction networks to verify the role of numerical weather forecast data and spatial bias characteristics in the day-ahead power generation prediction of photovoltaic power plants.

[0089] Apart from the input feature configuration and combined network structure, all experiments used the same data source, data preprocessing method, sampling interval, loss function and evaluation data.

[0090] During the testing process, the actual power generation during the predicted date is only used for comparison with the prediction results and is not used as input to the combined prediction model.

[0091] This embodiment employs three combined prediction networks: TCN-BiLSTM, TCN-LSTM, and CNN-LSTM.

[0092] I. Only historical power generation data is used. The first input configuration uses only historical power generation data as input to the prediction model, without incorporating numerical weather prediction data and spatial bias characteristics.

[0093] The prediction results are shown in Table 1.

[0094] Table 1. Prediction results of each model when only historical power generation data is used.

[0095] See Figure 4 When only historical power generation data is used, each model can reflect the basic intraday trend of photovoltaic power generation. However, during periods of high power such as midday when cloud cover is easily affected, there is a deviation between the predicted and actual values.

[0096] II. Incorporating Numerical Weather Prediction Data The second input configuration adds numerical weather forecast data to the historical power generation data.

[0097] See Figure 5 There are varying degrees of correlation between heterogeneous meteorological characteristics and photovoltaic power generation. The prediction results obtained after incorporating numerical weather prediction meteorological characteristics into the model input are shown in Table 2.

[0098] Table 2. Prediction results of each model after incorporating numerical weather prediction data.

[0099] Compared to the input configuration that only uses historical power generation data, after adding numerical weather prediction data, the root mean square error of TCN-BiLSTM, TCN-LSTM and CNN-LSTM decreased by 0.0241, 0.0188 and 0.0267 respectively, and the mean absolute error decreased by 0.0145, 0.0120 and 0.0203 respectively.

[0100] See Figure 6Under clear weather conditions, the prediction curves before and after adding numerical weather forecast data can reflect the overall power generation trend; under cloudy and rainy weather conditions, the prediction curves after adding numerical weather forecast data can better track the power generation fluctuations caused by meteorological changes.

[0101] III. Incorporating Spatial Bias Features The third input configuration, based on the second input configuration, further incorporates temperature space deviation, global horizontal irradiance space deviation, and direct irradiance space deviation.

[0102] The three types of spatial biases are calculated according to formulas (4) to (6) in Example 1, and the three types of historical spatial biases are combined with historical power generation and the numerical weather forecast meteorological characteristics of the daytime to be predicted according to formula (6-1).

[0103] The prediction results are shown in Table 3.

[0104] Table 3. Prediction results of each model after incorporating spatial bias features.

[0105] After incorporating spatial bias characteristics, the root mean square error of TCN-BiLSTM decreased from 0.0142 to 0.0113, the mean absolute error decreased from 0.0096 to 0.0085, and the mean error decreased from 0.0005 to 0.0001.

[0106] The correlation coefficient of the TCN-BiLSTM model changed from 0.9987 to 0.9938. Therefore, this embodiment does not summarize the improvement of the correlation coefficient as obtained after adding spatial bias features.

[0107] See Figure 7 Under clear weather conditions, the prediction curves before and after adding spatial bias features can reflect the overall trend of actual power generation. Under cloudy and rainy weather conditions, the deviation between the prediction curve and the actual power generation curve is reduced after adding spatial bias features.

[0108] The experimental results above show that numerical weather prediction data is used to provide macro-meteorological information for the forecast period before the actual date. The spatial deviations of temperature, global horizontal irradiance, and direct irradiance formed before the forecast release time are used to characterize the historical local deviations of the target photovoltaic power station relative to the regional average meteorological conditions. Inputting historical power, historical local deviation information, and numerical weather prediction information for the forecast period before the actual date into the TCN-BiLSTM combined prediction model can reduce the forecast error of day-ahead power generation under complex weather conditions.

[0109] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, equipment, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0110] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0111] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program instructions, such as USB flash drives, portable hard drives, read-only storage servers, random access storage servers, magnetic disks, or optical disks.

[0112] Furthermore, it should be noted that the combination of the various technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0113] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

[0114] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM, executed by a power prediction device, comprising: Historical sample data is obtained, including historical power generation data of the target photovoltaic power station, historical surface meteorological observation data of the target photovoltaic power station and multiple reference photovoltaic power stations in its area, and numerical weather forecast data corresponding to the historical day-ahead period; The regional meteorological background is determined based on the historical surface meteorological observation data of the target photovoltaic power station and the multiple reference photovoltaic power stations. The spatial deviation of temperature, the spatial deviation of global horizontal irradiance, and the spatial deviation of direct irradiance are determined based on the differences between the historical surface meteorological observation values ​​of the target photovoltaic power station and the corresponding regional meteorological background. The historical power generation sequence and the spatial deviation sequence located within the historical input period are combined with the numerical weather forecast sequence of the corresponding historical day-ahead period to form historical training samples. The actual power generation sequence of the historical day-ahead period is used as the label to train a combined prediction model including a temporal convolutional network, a bidirectional long short-term memory network and a fully connected output layer. When executing the forecast, obtain the historical power generation sequence and historical spatial deviation sequence of the target photovoltaic power station before the forecast release time, and obtain the numerical weather forecast sequence corresponding to the forecast day; According to the same feature order as the historical training samples, the historical power generation sequence, the historical spatial deviation sequence, and the numerical weather forecast sequence corresponding to the prediction day are combined into a multi-source input sample to be predicted; The combined prediction model, trained by inputting the multi-source input samples to be predicted, extracts local temporal variation features through the temporal convolutional network, extracts bidirectional temporal dependency features through the bidirectional long short-term memory network, and outputs the day-ahead power generation prediction sequence of the target photovoltaic power station through the fully connected output layer.

2. The day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM as described in claim 1, wherein, The historical power generation data, historical surface meteorological observation data, and numerical weather forecast data are preprocessed, including: removing data from periods without sunshine and retaining valid sunshine period data from 6:00 to 21:00 each day; matching the data according to valid timestamps; and normalizing the data in different dimensions using the maximum-minimum normalization method.

3. The day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM according to claim 2, characterized in that, The temporal clarity index is calculated based on the predicted total horizontal irradiance from historical numerical weather forecast data and the historical measured global horizontal irradiance at the corresponding time. A first preset weather classification threshold and a second preset weather classification threshold are set, wherein the first preset weather classification threshold is greater than the second preset weather classification threshold. When the temporal clarity index is greater than or equal to the first preset weather classification threshold, the corresponding historical time is classified as a sunny weather condition. When the temporal clarity index is greater than or equal to the second preset weather classification threshold and less than the first preset weather classification threshold, the corresponding historical time is classified as a cloudy weather condition. When the temporal clarity index is less than the second preset weather classification threshold, the corresponding historical time is classified as a rainy weather condition.

4. The day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM as described in claim 3, wherein, Each historical moment and test moment is assigned to a corresponding weather condition subset according to the aforementioned weather conditions. The weather condition subsets are used to evaluate the prediction performance of the combined prediction model under sunny, cloudy, and rainy conditions, respectively.

5. The day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM according to claim 1, wherein, The regional meteorological background includes the average temperature, average global horizontal irradiance, and average direct irradiance of the target photovoltaic power station and the multiple reference photovoltaic power stations at corresponding historical times; the spatial deviation of temperature is the difference between the historical measured temperature of the target photovoltaic power station and the average temperature, the spatial deviation of global horizontal irradiance is the difference between the historical measured global horizontal irradiance of the target photovoltaic power station and the average global horizontal irradiance, and the spatial deviation of direct irradiance is the difference between the historical measured direct irradiance of the target photovoltaic power station and the average direct irradiance.

6. The day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM according to claim 5, wherein, The historical input period is located before the corresponding historical day-ahead period; the historical input period and the historical day-ahead period have the same sampling interval and the same number of sampling points, and the historical power generation sequence, the spatial deviation sequence and the numerical weather forecast sequence are combined according to the same sampling sequence number; when performing day-ahead forecasting, the time end of the historical power generation sequence and the historical spatial deviation sequence is not later than the forecast release time.

7. The day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM according to claim 6, wherein, The bidirectional long short-term memory network includes forward long short-term memory units and reverse long short-term memory units. The forward long short-term memory units process the features output by the temporal convolutional network in forward time order of the input sequence, and the reverse long short-term memory units process the features output by the temporal convolutional network in reverse time order of the input sequence, and concatenate the forward hidden state and the reverse hidden state at the same time step; the fully connected output layer outputs the day-ahead normalized power generation prediction value based on the concatenated bidirectional hidden state.

8. The day-ahead prediction method for photovoltaic power generation based on spatial downscaling and TCN-BiLSTM according to claim 1, wherein, The hyperparameters of the combined prediction model are optimized using a Bayesian optimization algorithm. The hyperparameters include the number of network channels of the temporal convolutional network, the number of hidden layer neurons of the bidirectional long short-term memory network, the initial learning rate, the residual block dropout rate, and the batch sample processing number.

9. A method for day-ahead prediction of photovoltaic power generation based on spatial downscaling and TCN-BiLSTM according to claim 8, wherein, The root mean square error is used as the loss function of the combined prediction model, and the network parameters of the combined prediction model are updated using an adaptive moment estimation optimizer.

10. A method for predicting the day-ahead power generation of a photovoltaic power plant based on spatial downscaling and TCN-BiLSTM according to claim 9, wherein, The temporal convolutional network has 32 network channels, the bidirectional long short-term memory network has 64 hidden layer neurons, the initial learning rate is 0.0024, the residual block dropout rate is 0.3, and the batch sample processing number is 128.

Images