Method and system for probabilistic forecasting of next day generation sequence of multiple renewable energy power plants

Abstract

A method of making next-day power generation sequence probability forecasts for a plurality of renewable energy power plants. The method comprises the steps of: a) for each client device, mapping its raw data inputs into latent features; b) transmitting to a server a prediction model that each client device locally hosts in the form of latent features and model parameters; c) aggregating the localised prediction models of the plurality of client devices at the server side; d) distributing the aggregated model to the client devices; e) updating the localised prediction models of the client devices based on the aggregated model; f) generating power output sequence probability forecasts on each client device based on the updated localised prediction models. The plurality of client devices respectively correspond to specific facilities in the plurality of renewable energy power plants. All client devices are connected to the server.

Description

Technical Field

[0001] This invention relates to renewable energy power plants, and in particular to probabilistic prediction of renewable energy power plants. Background Technology

[0002] Renewable energy sources such as wind and solar power are crucial for achieving carbon neutrality, and their global development continues to be rapid [1]. The importance of studying renewable energy generation forecasting (RPF) methods to address generation uncertainty in grid operation has been widely recognized [2]. Accurate and reliable renewable energy generation forecasts are crucial for grid operation and can help grid operators better predict sudden changes in generation caused by renewable energy. For example, as described in reference [3], improved forecasting techniques can significantly reduce ramp rate violations. Based on the forecast timeframe, RPF tasks can be broadly categorized into short-term (0-6 hours in advance), medium-term (6-24 hours in advance), and long-term (more than 24 hours in advance) forecasting tasks [4]. Methods with different forecast timeframes have different practical value. RPF results covering longer future periods can support downstream tasks such as electricity pricing [5], unit dispatching [6], energy storage management [7], power plant maintenance [8], and electricity trading [9]. However, the accumulation of uncertainty in renewable energy supply within the forecast timeframe makes this task extremely challenging.

[0003] The core issue of RPF research is the development of predictive models. The widespread deployment of Supervisory Control and Data Acquisition (SCADA) systems in commercial renewable energy power plants has provided an unprecedented opportunity for RPF research to transform from traditional physical models

[10] and statistical models

[11] to today's data-driven models, including classical machine learning models

[12] and cutting-edge deep learning models

[13] . Based on the output type, RPF tasks can be divided into deterministic and probabilistic categories. Deterministic RPF aims to predict future spot value and has been extensively studied in the literature. In contrast, probabilistic RPF quantifies the uncertainty of future renewable energy generation by providing confidence intervals, quantiles or distributions of the prediction. Since probabilistic RPF has the ability to quantify the uncertainty of renewable energy generation, its generated results provide richer information for risk assessment and operation decisions of power systems

[14] .

[0004] The core of probabilistic RPF research lies in exploring various methods to improve probabilistic prediction performance. Parametric and nonparametric probabilistic RPF methods have been extensively discussed. Parametric methods usually specify the form of the prediction density function (such as Gaussian or Beta distribution) and rely on only a few parameters. In

[15] , the Bayesian Information Criterion (BIC) is used to select the optimal parameters for the sparse vector autoregression algorithm for RPF. In

[16] , an improved Taylor Kriging method is developed to estimate the parameters of future renewable energy generation density. On the other hand, nonparametric methods quantify the uncertainty of renewable energy generation by estimating confidence intervals, quartiles and distributions without making any assumptions about the shape of the distribution. Quantile regression (QR) [17,18] and upper and lower bound estimation (LUBE) [19,20] are applied to probabilistic RPF research by directly estimating confidence intervals (CI). Reference

[17] proposes a temporal convolution QR method that combines QR with temporal convolutional networks (TCN) for wind power quantile estimation. In

[19] , a neural network-based LUBE and moving block bootstrapping method were proposed for probabilistic RPF. In

[20] , a LUBE combined with a recurrent neural network (RNN) was developed to enhance RPF. In addition to prediction intervals, kernel density estimation (KDE) methods [21,22] were developed to infer the distribution of future power generation and realize RPF. In

[21] , the bivariate vector autoregressive moving average-generalized autoregressive conditional heteroscedasticity method was applied to RPF. In

[22] , a KDE model with four bandwidth selectors was introduced for RPF. Recently, hybrid density networks (MDN) [23-29] have attracted attention as a more flexible paradigm for developing probabilistic prediction models. Compared with KDE, MDN uses a multi-component hybrid model instead of a single-kernel distribution, thereby achieving greater modeling flexibility and more accurate prediction. Research

[23] integrated convolutional neural networks (CNNs) with recurrent neural units (GRUs) to form an MDN inference network. Due to its simplicity and ease of use, as well as its suitability for sampling and likelihood calculation, Gaussian mixture models (GMMs) were used as the probability density function (PDF). However, GMMs can lead to density leakage in the mixture model. Research

[24] addressed this problem by replacing GMMs with beta kernel MDNs. Research

[25] further improved MDN performance by employing an adversarial learning algorithm based on Wasserstein distance to train the model.

[0005] Most previous probabilistic RPF studies [15-29] have focused on developing more sophisticated methods to improve prediction performance, ideally assuming perfect data availability. Meanwhile, to develop more predictive features, a few recent studies [4, 32, 33] have begun to utilize a wider range of data sources. As shown in Table 1, based on recent literature [31-36], two main directions of RPF research can be clearly observed: 1) constructing network architectures for long-term RPF tasks based on target renewable energy data; and 2) designing network architectures that utilize information from multiple data sources. However, the assumption of perfect data availability in previous studies may pose a significant burden in real-world applications involving data privacy and security issues

[43] . Furthermore, due to data sharing regulations, data acquisition from distributed renewable energy units across different power plants in different regions may not be feasible.

[0006] To address privacy concerns in data-driven modeling, the federated learning (FL) paradigm [44-46] has emerged. This paradigm employs a client-server architecture, comprising multiple clients (for collecting local data) corresponding to a power plant and a server (for aggregating desensitized data and models). Recent literature has developed various FL-based methods for privacy-preserving RPF tasks [37-42, 44, 46, 49]. In

[44] , the general FedAvg framework was used to integrate multi-source data without compromising privacy. In

[49] , a personalized federated learning (PFL) strategy was employed to enhance the system's robustness to anomalous updates from a single wind farm. Based on observations in [37-42], research on FL-based RPF methods primarily focuses on privacy protection in deterministic RPF problems, while research on probabilistic RPF methods that also consider privacy is relatively scarce. Furthermore, most studies [37-42, 49] only address short-term RPF tasks. In the field of renewable energy forecasting, day-ahead series forecasting has more comprehensive characteristics than short-term forecasting—its forecast series has covered the time periods considered in many short-term forecasting studies [30-33]. At the same time, incorporating data privacy protection mechanisms into the development of probabilistic series forecasting models has value-added service significance. Since day-ahead forecasting plays a key role in operation planning, energy trading and market participation, it can not only ensure local data privacy and security, but also realize forecasting methods with richer information modeling, which can not only improve modeling efficiency, but also avoid unfair competition caused by data leakage and information abuse

[50] . Therefore, from a practical point of view, it is more valuable to study privacy protection modeling in probabilistic renewable energy next-day series forecasting (DRPSF).

[0007] References All the following references (and related appendices and / or supplementary materials) are incorporated in full into this application by way of citation: [1]S. Letzgus S, K. R. Müller. An explainable AI framework for robustand transparent data-driven wind turbine power curve models. Energy and AI,2024, 15: 100328. [2]H. Chen, A novel wind model downscaling with statisticalregression and forecast for the cleaner energy. Journal of Cleaner Production434 (2024): 140217. [3]X. Peng, Y. Li, F. Tsung. A Graph Attention Network with Spatio-Temporal Wind Propagation Graph for Wind Power Ramp Events Prediction.Renewable Energy 236 (2024) 121280. [4]M. Khodayar, J. Wang. Spatio-temporal graph deep neural networkfor short-term wind speed forecasting. IEEE Transactions on SustainableEnergy 10(2) (2018) 670-681. [5]J. Trebbien, L. R. Gorjão, A. Praktiknjo, et al. Understandingelectricity prices beyond the merit order principle using explainable AI.Energy and AI, 2023, 13: 100250. [6]J. Wang, M. Shahidehpour, Z. Li. Security-constrained unitcommitment with volatile wind power generation.IEEE Transactions on PowerSystems 23(3) (2008) 1319–1327. [7]R. Blonbou, S. Monjoly, J. F. Dorville. An adaptive short-termprediction scheme for wind energy storage management. Energy conversion andmanagement 52(6) (2011) 2412–2416. [8]H. Sharma H, L. Marinovici, V. Adetola, et al. Data-drivenmodeling of power generation for a coal power plant under cycling. Energy andAI, 2023, 11: 100214. [9]A. Pircalabu, T. Hvolby, J. Jung, E. Høg. Joint price andvolumetric risk in wind power trading: A copula approach. Energy Economics 62(2017) 139–154.

[10] U. Focken, M. Lange, H. P. Waldl, Previento-a wind powerprediction system with an innovative upscaling algorithm in Proceedings ofthe European Wind Energy Conference, Copenhagen, Denmark 276 (2001).

[11] X. Liu, Z. Lin, Z. Feng. Short-term offshore wind speed forecastby seasonal ARIMA-A comparison against GRU and LSTM. Energy 227 (2021)120492.

[12] S. Wang, B. Li B, G. Li, et al. A Comprehensive Review on theDevelopment of Data-Driven Methods for Wind Power Prediction and AGCPerformance Evaluation in Wind-Thermal Bundled Power Systems. Energy and AI,2024: 100336.

[13] J. Zhu, L. Su, Y. Li, Wind power forecasting based on new hybridmodel with TCN residual modification. Energy and AI, 2022, 10: 100199.

[14] Z. Zheng, Z. Zhang, A Stochastic Recurrent Encoder DecoderNetwork for Multistep Probabilistic Wind Power Predictions, IEEE Transactionson Neural Networks and Learning Systems (2023).

[15] J. Dowell, P. Pinson, Very-Short-Term Probabilistic Wind PowerForecasts by Sparse Vector Auto-regression, IEEE Transactions on Smart Grid 7(2015) 736–770.

[16] H. Liu, J. Shi, E. Erdem, Prediction of wind speed time seriesusing modified Taylor Kriging method, Energy 35 (2010) 4870–4879.

[17] J. Hu, Q. Luo, J. Heng, Y. Deng, Conformalized temporalconvolutional quantile regression networks for wind power intervalforecasting, Energy 248 (2022) 123497.

[18] J. Hu, J. Tang, Y. Lin. A novel wind power probabilisticforecasting approach based on joint quantile regression and multi-objectiveoptimization. Renewable Energy 149 (2020) 141-164.

[19] A. Khosravi, S. Nahavandi, D. Creighton, Prediction intervals forshort-term wind farm power gen-eration forecasts, IEEE Transactions onsustainable energy 5 (2013) 602–610.

[20] Z. Shi Z, H. Liang, V. Dinavahi. Direct interval forecast ofuncertain wind power based on recurrent neural networks, IEEE Transactions onsustainable energy 9 (2017) 1177–1187.

[21] J. Jeon, J. W. Taylor, Using conditional kernel densityestimation for wind power density forecasting, Journal of the AmericanStatistical Association 107 (2012) 66–79.

[22] Q. Han, S. Ma, T. Wang, F. Chu, Kernel density estimation modelfor wind speed probability distribution with applicability to wind energyassessment in China, Renewable and Sustainable Energy Reviews, 115 (2019)109387.

[23] M. Afrasiabi, M. Mohammadi, M. Rastegar, S. Afrasiabi Advanceddeep learning approach for probabilistic wind speed forecasting, IEEETransactions on Industrial Informatics 17 (2020) 720–727.

[24] Z. Men, E. Yee, F. S. Lien, et al. Short-term wind speed andpower forecasting using an ensemble of mixture density neural networks.Renewable Energy 87 (2016) 203-211.

[25] L.Yang, Z. Zheng, Z. Zhang, An Improved Mixture Density NetworkVia Wasserstein Distance Based Adversarial Learning for Probabilistic WindSpeed Predictions, IEEE Transactions on Sustainable Energy 13, (2021) 755–766.

[26] A. Brusaferri, M. Matteucci, S. Spinelli, et al. Probabilisticelectric load forecasting through Bayesian mixture density networks. AppliedEnergy 309 (2022) 118341.

[27] R. Raidoo, R. Laubscher, Data-driven forecasting with modeluncertainty of utility-scale air-cooled condenser performance using ensembleencoder-decoder mixture-density recurrent neural networks, Energy 238 (2022)122030.

[28] Z. Zheng, Z. Zhang. A stochastic recurrent encoder decodernetwork for multistep probabilistic wind power predictions. IEEE Transactionson Neural Networks and Learning Systems, in press, 2023.

[29] H. Xue, Y. Jia, P. Wen, G. Saeid, Using of improved models ofGaussian Processes in order to Regional wind power forecasting, Journal ofClean Production 262 (2020) 121391.

[30] M. Yang, D. Wang, W. Zhang. A short-term wind power predictionmethod based on dynamic and static feature fusion mining. Energy 280 (2023)128226.

[31] M. A. Hossain, E. Gray, J. Lu, et al. Optimized forecasting modelto improve the accuracy of very short-term wind power prediction. IEEETransactions on Industrial Informatics 19(10) (2019) 10145-10159.

[32] H. Liu H, L. Yang, B. Zhang, et al. A two-channel deep networkbased model for improving ultra-short-term prediction of wind power viautilizing multi-source data. Energy 283 (2023) 128510.

[33] C. A. Severiano, P. C. L. e Silva, M. W. Cohen, et al. Evolvingfuzzy time series for spatio-temporal forecasting in renewable energysystems. Renewable Energy 171 (2021) 764-783.

[34] C. Wan, Z. Xu, P. Pinson, et al. Probabilistic forecasting ofwind power generation using extreme learning machine. IEEE Transactions onPower Systems 29(3) (2013) 1033-1044.

[35] J. J. Mesa-Jiménez, A. L. Tzianoumis , L. Stokes, et al. Long-term wind and solar energy generation forecasts, and optimisation of PowerPurchase Agreements. Energy Reports 9 (2023) 292-302.

[36] A. Ahmadi A, M. Nabipour, B. Mohammadi-Ivatloo, et al. Long-termwind power forecasting using tree-based learning algorithms. IEEE Access 8(2020) 151511-151522.

[37] C. Goncalves, J. B. Ricardo, and P Pinson. "Privacy-preservingdistributed learning for renewable energy forecasting." IEEE Transactions onSustainable Energy 12.3 (2021): 1777-1787.

[38] Y. Li, et al. "Wind power forecasting considering data privacyprotection: A federated deep reinforcement learning approach." Applied Energy329 (2023): 120291.

[39] X. Zhang, F Fang,J Wang. "Probabilistic solar irradiationforecasting based on variational Bayesian inference with secure federatedlearning." IEEE Transactions on Industrial Informatics 17.11 (2020): 7849-7859.

[40] L. Zhang, S. Zhu, S. Su, et al. Wind power prediction methodbased on cloud computing and data privacy protection. Journal of CloudComputing, 2024, 13(1): 1-14.

[41] Y. Wang Y, Q. Guo. Privacy-Preserving and Adaptive Federated DeepLearning for Multiparty Wind Power Forecasting. IEEE Transactions on IndustryApplications, 2024.

[42] A. Alshardan, S. Tariq, R. N. Bashir, et al. Federated Learning(FL) Model of Wind Power Prediction. IEEE Access, 2024.

[43] Y. Li, J. Li, Y. Wang, Privacy-preserving spatiotemporal scenariogeneration of renewable energies: A federated deep generative learningapproach, IEEE Transactions on Industrial Informatics (2021).

[44] K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman, V.Ivanov, et al, Towards Federated Learning at Scale: System Design, inproceedings of Machine Learning and Systems 1 (2019) 374–388.

[45] T. Li, A. K. Sahu, A. Talwalkar, Federated learning: Challenges,methods, and future directions, IEEE Signal Processing Magazine 37 (2020) 50–60.

[46] L. Li, Y. Fan, M. Tse, K. Y. Lin, A review of applications infederated learning, Computers&Industrial Engineering 149 (2020) 106854.

[47] X. Zhao, B. Sun, and R. Geng, A new distributed decomposition–reconstruction–ensemble learning para-digm for short-term wind powerprediction. Journal of Cleaner Production 423 (2023) 138676.

[48] Y. Qiang, Y. Liu, T. Chen, Y. Tong. Federated Machine Learning:Concept and Applications, ACM Transactions on Intelligent Systems andTechnology 10 (2019) 1–19.

[49] Y. Zhao, S. Pan, Y. Zhao, et al. Ultra-short-term wind powerforecasting based on personalized robust federated learning with spatial collaboration. Energy 288 (2024) 129847.

[50] Y. Li, R. Wang, Y. Li, M. Zhang, C. Long, Wind power forecasting considering data privacy protection: A federated deep reinforcement learning approach, Applied Energy 329 (2023) 120291.

[51] H. Liu and Z. Zhang. Development and Trending of Deep LearningMethods for Wind Power Predictions. Artificial Intelligence Review 57 (2024)112. Summary of the Invention

[0008] Based on the above background, one of the objectives of this invention is to use a more advanced modeling method to study the task of renewable energy power generation prediction, that is, to make probabilistic predictions of the day-ahead power generation sequence of multiple renewable energy power plants without infringing on the data privacy of each power plant.

[0009] The present invention provides a computer-implemented method for probabilistically predicting the next-day power generation sequence of multiple renewable energy power plants while protecting data privacy. The method includes the following steps: a) mapping raw data input to latent features for each client device, wherein multiple client devices correspond to each of the multiple renewable energy power plants; b) transmitting a locally hosted prediction model for each client device, in the form of the latent features of the client device and model parameters, to a server, wherein multiple client devices are connected to the server; c) aggregating the locally hosted prediction models of the multiple client devices at the server; d) distributing the aggregated model to the multiple client devices; e) updating the localized prediction model on each of the multiple client devices based on the aggregated model; and f) generating a probabilistic prediction of the power output sequence on each of the multiple client devices based on the updated localized prediction model.

[0010] Preferably, for each of the multiple client devices, step a) is performed by a local feature extractor on that client device.

[0011] More preferably, the local feature extractor is a deep neural network (DNN), a convolutional neural network (CNN), a long short-term memory network (LSTM), or a gated recurrent unit (GRU).

[0012] In a variation of the preferred embodiment, the local feature extractor is assisted by a discriminator on the server to identify domain-invariant features.

[0013] In another variation of the preferred embodiment, the latent feature is a domain-invariant feature.

[0014] In yet another variation of the preferred embodiment, the model parameters are generated by a local probability estimator of each of the plurality of client devices.

[0015] In yet another variation of the preferred embodiment, the server includes a global feature extractor, a global probability estimator, and a discriminator.

[0016] Preferably, the global feature extractor is adapted to aggregate all latent features from multiple client devices. The global probability estimator is adapted to aggregate all model parameters from multiple client devices.

[0017] Preferably, the aggregation model includes aggregated latent features and aggregated model parameters, used to update local feature extractors and local probability estimators on multiple client devices.

[0018] Preferably, the discriminator is adapted to classify the domain labels of potential features.

[0019] In another variation of the preferred embodiment, steps a)-e) of the above method are repeated through multiple iterations to train a locally hosted prediction model.

[0020] In another variation of the preferred embodiment, the above method further includes the following steps: g) training a local probability estimator on at least one client device using a training dataset and a validation dataset to maximize the log likelihood of the probability predictions made by the local probability estimator.

[0021] In another variation of the preferred embodiment, the above method further includes the following steps: h) training a discriminator using features generated by multiple local feature extractors located on multiple client devices in different domains to maximize the log-likelihood value of the predicted label being equal to the domain label.

[0022] In another variation of the preferred embodiment, the above method further includes: training a local feature extractor on at least one client device using a combination loss derived from training a local probability estimator on at least one client device and a discriminator on a training server.

[0023] According to another aspect of the present invention, a system is provided for probabilistically predicting the next-day power generation sequence of multiple renewable energy power plants. The system includes one or more processors and a memory storing one or more programs executed by the one or more processors. The one or more programs contain instructions adapted to perform the computer-implemented method or variations thereof for probabilistically predicting the next-day power generation sequence of multiple renewable energy power plants described above.

[0024] According to another aspect of the present invention, a system for probabilistically predicting the next-day power generation sequence of multiple renewable energy power plants is provided. The system includes a server and multiple client devices connected to the server. The server is adapted to aggregate locally hosted prediction models from the multiple client devices and distribute the aggregated models to the multiple client devices. The locally hosted prediction models received by the server include the potential features and model parameters of these locally hosted prediction models.

[0025] According to another aspect of the invention, a non-transitory computer-readable storage medium is provided, which stores one or more programs suitable for execution by one or more processors. The one or more programs contain instructions suitable for performing the above-described computer-implemented method for probabilistically predicting the next-day power generation sequence of multiple renewable energy power plants, or variations thereof.

[0026] According to another aspect of the present invention, a method for probabilistically predicting the next-day power generation sequence of multiple renewable energy power plants is provided without infringing on the data privacy of each power plant. The method includes: implementing a server-client based system to coordinate network models to simultaneously predict the next-day power generation sequence of wind turbines and solar panels of multiple renewable energy power plants; extracting features through client-side decoding, and hosting a network-based local feature extractor and a network-based probabilistic predictor; maintaining global versions of the feature extractor and the probabilistic predictor on the server side, and developing a discriminative network.

[0027] In some embodiments, the "server-client" architecture consists of multiple local networks distributed across renewable energy power plants and a global network hosted on cloud servers.

[0028] In some embodiments, the overall framework is built on a deep network model.

[0029] In some embodiments, the local probability estimator in the client architecture employs a hybrid density network (MDN), which is a type of fully connected neural network.

[0030] In some embodiments, to extract latent features from local data, the client may select one of four widely used deep networks: fully connected neural networks (DNN), convolutional neural networks (CNN), long short-term memory networks (LSTM), and circular update units (GRU).

[0031] In some embodiments, the server consists of three collaborative networks: a global feature extractor, a global probability estimator, and a discriminator.

[0032] In some embodiments, each iteration of the network execution process includes four steps: forward propagation, model aggregation, back propagation, and model output.

[0033] In some embodiments, the forward propagation step extracts local latent features using a local feature extractor.

[0034] In some embodiments, the forward propagation step decodes local latent features through a local probability estimator to generate probability predictions.

[0035] In some embodiments, the discriminator on the server side assists the local feature extractor in identifying domain-invariant features by classifying local features by domain labels during the forward propagation step.

[0036] In some embodiments, the server-side global feature extractor and probability estimator aggregate knowledge from the local model by averaging.

[0037] In some embodiments, the network parameters serving the local feature extractor and the local probability estimator are updated during the backpropagation step based on the combined loss of probability prediction and prediction domain labels.

[0038] In some embodiments, the local feature extractor and probability estimator use values ​​distributed by the server.

[0039] In some embodiments, the local feature extractor is trained using a combination loss.

[0040] In some embodiments, the discriminator is trained using features generated by different domain feature extractors to maximize the log-likelihood that the predicted label is equal to the domain label.

[0041] In some embodiments, the discriminator attempts to distinguish domain labels of local features.

[0042] Different exemplary embodiments of the present invention thus provide an advanced Domain Invariant Feature Learning Embedded Joint Learning (DIFL) framework, which consists of multiple client networks distributed across renewable energy power plants and a global network hosted on a cloud server. Each client extracts features by decoding, and respectively hosts a network-based local feature extractor and a network-based probabilistic predictor. The two networks are used as the backbone network by considering the server-side domain invariant feature extractor and global probabilistic predictor. The server maintains global versions of the feature extractor and probabilistic predictor and constructs a discriminator network to perform a two-stage process: 1) aggregating the received local models to develop the global feature extractor and probabilistic predictor; 2) training the global feature extractor through the discriminator, enabling the local feature extractor adapted to the global model to achieve stronger robustness in local latent feature engineering.

[0043] Given its privacy-preserving features and predictive performance, this invention is not limited to the field of long-term probabilistic renewable energy sequence prediction.

[0044] The above summary is for illustrative purposes only and is not restrictive. This disclosure includes one or more corresponding aspects, embodiments, or features, whether used alone or in various combinations, whether expressly stated in the combination or when used alone (including in the claims). It should be understood that a feature defined according to any aspect of this disclosure, or a feature relating to any specific embodiment of this disclosure, may be used alone or in combination with any other defined feature, used in any other aspect or embodiment, or to form a further aspect or embodiment of this disclosure. Attached Figure Description

[0045] The above and more features of the present invention will be illustrated by the following embodiments, which are merely illustrative and are described in conjunction with the accompanying drawings, wherein: Figure 1 The prediction processes of the classic RPF and the privacy-preserving RPF are demonstrated.

[0046] Figure 2 This is a schematic diagram of the DIFL method according to the first embodiment of the present invention.

[0047] Figure 3 A local probability estimator is presented.

[0048] Figure 4 The client device was presented. i The training and testing process.

[0049] Figure 5 This demonstrates the random selection of time periods for DIFL's prediction intervals on WPD.

[0050] Figure 6 This demonstrates the random selection of time periods for the DIFL prediction interval on the SPD.

[0051] Figure 7 The correlation analysis between PICP and PINAW is presented.

[0052] Figure 8 illustrates an exemplary information processing apparatus structure that can be used to implement the above method. Detailed Implementation

[0053] An exemplary embodiment of the present invention provides an advanced DIFL framework for coordinating the development of a deep network model-based system comprising multiple clients and a server. In DIFL, each client serving a local renewable energy power plant maps raw data input to latent features through a local feature extractor and generates probabilistic predictions of power output sequences using a locally located prediction model. The cloud-based server first aggregates the knowledge from each client's model and then distributes the aggregated model back to each client, allowing each local feature extractor to identify domain-invariant features by interacting with a server-side discriminator. Therefore, only desensitized data such as model parameters are allowed to be transmitted, ensuring the local data privacy of the power plant. To verify the advantages of DIFL, its theoretical properties are first explored. Subsequently, computational studies are conducted based on datasets collected from commercial renewable energy power plants, comparing DIFL with well-known benchmark models. The results further confirm that regardless of whether real wind farm or solar power plant datasets are used, DIFL consistently outperforms all benchmark models on average.

[0054] Table 1: Comparison of Key Recent Advances in Data-Driven RPF Research

[0055] To address more challenging tasks, such as the privacy-preserving DRPSF (Distributed Probabilistic Prediction System), the advanced DIFL (Domain Invariant Feature Learning) framework consists of multiple client networks distributed across renewable energy power plants and a global network of servers in the cloud. Each client extracts features by decoding, running both a network-based local feature extractor and a network-based probabilistic predictor. The server-side domain invariant feature extractor and global probabilistic predictor form the backbone network. The server maintains global versions of the feature extractor and probabilistic predictor and constructs a discriminator network to perform the following two-phase process: 1) aggregating received local models to develop the global feature extractor and probabilistic predictor; 2) training the global feature extractor through the discriminator, helping the local feature extractor adapted from the global model to achieve stronger robustness in local latent feature engineering. Iterative interactions between clients and the server primarily involve transferring local models to the server and distributing the updated global model back to each client. DIFL achieves knowledge transfer between clients without leaking local data, as the information transferred between clients and the server only contains latent features and model parameters. DIFL effectively reduces server load by distributing computational workload between the client and server. To verify its advantages, the study mathematically proves that the prediction error in the target domain can be jointly limited by the source domain modeling quality and the divergence between the source and target domains. Furthermore, a comprehensive computational study based on a commercial renewable energy power plant dataset, benchmarking DIFL against well-known benchmark models, shows that DIFL outperforms the classic RPF framework trained on complete data in most cases. This method contributes to existing technologies in four aspects: Privacy Protection and More Advanced RPF Tasks: This paper proposes a more advanced RPF task with privacy protection features—multi-factory probabilistic DRPSF. As shown in Table 1, existing privacy-preserving RPF model research [37-42] mainly targets deterministic RPF problems and short-term prediction tasks; Privacy-Preserving Converged DRPSF: This study proposes a coordination system based on a "server-client" architecture to synchronously predict the next-day power generation sequences of wind turbines and solar panels from multiple renewable energy power plants through a network model. As shown in Table 1, existing privacy-preserving RPF studies [37-42] mostly handle wind power and solar power prediction tasks separately; A Novel Modeling Pipeline Based on Domain Invariant Learning: This paper proposes a novel DIFL method to construct a more efficient probabilistic RPF model for each renewable energy power plant by leveraging group-level knowledge (while ensuring local data privacy). Unlike previous FL studies [37-42, 49], this method develops a discriminator on the server side to approximate the upper bound of the differences in the local data feature space, thereby helping the local feature extractor to obtain domain invariant features. This method only transmits latent features and model parameters between the client and the server, thus ensuring data privacy. To the inventors' knowledge, DIFL is the first research to integrate invariant feature engineering from different domains into a privacy-preserving task. Superior Modeling Performance: Based on the comprehensive computational experiments conducted in this study, DIFL constructs a robust probabilistic RPF model for each renewable energy power plant by incorporating privacy-preserving group knowledge. Furthermore, an upper bound on the prediction error of DIFL is derived through mathematical analysis.

[0056] The following section first outlines the classic next-day probabilistic RPF modeling method that uses only data from a single power plant (such as...). Figure 1 (As shown on the left), followed by an introduction to the RPF modeling scheme proposed in this study that is based on multi-site data and also takes privacy into account (e.g.) Figure 1 (As shown on the right).

[0057] set up Indicates containing from N Historical power generation collected from different wind or solar power plants M Multivariate time series Indicates from the first indivual( Input records collected by the power plant. The classic probabilistic RPF task aims to build a data-driven model to estimate the probabilistic prediction of the next day's electricity output, which can be expressed as quantiles, upper and lower limits, or probability distributions. In practice, a mixture conditional density function (such as...) is typically used to describe it. Figure 1 (As shown on the left). The specific modeling framework is given by formulas (1) and (2):

[0058] in This indicates the number of components in the mixture model. Satisfies formula (3), It has parameters The distribution of beta is usually chosen in practice, as described by formula (4). As a component of the hybrid model.

[0059]

[0060] in For the gamma function. To select the... i Optimal parameters for a power plant and The maximum likelihood estimation (MLE) method is typically used to maximize the dataset based on equation (5). Actual power generation The log-likelihood value.

[0061]

[0062] This section proposes an extension to the classic RPF model, achieving NWP integration by constructing a privacy-preserving modeling framework. N Historical data of individual power plants Efficiently utilize information (such as) Figure 1 (As shown on the right). To integrate data from different power plants while protecting privacy, a Privacy Preservation Modeling (PPM) paradigm was designed. This paradigm first decomposes the data-driven model into two modules: a feature extractor... A probability estimator used to obtain local latent features. Used to provide the next day's renewable energy generation sequence The distribution of, among which The predicted sequence length is shown in equations (6)–(8).

[0063]

[0064] This modeling process allows for the transfer of desensitized information, including latent features. as well as The module's transmission, thereby improving system performance through the following two technologies: First, a global feature extractor was designed to integrate knowledge learned from different clients. With global probability estimator The aggregations obtained through formulas (9) and (10) are respectively from and The integrated model was then distributed back to the respective clients.

[0065]

[0066] Secondly, a discriminator was designed to help the feature extractor generate domain-invariant features. To predict local features domain tags .

[0067] Based on this modeling architecture, the local model can integrate knowledge from multiple sources and improve prediction performance. However, it was also found that local predictions based on the fusion of knowledge from other sources have performance boundaries. Let... Represents the target domain i Data distribution This represents the data distribution of the source domain (the concatenation of all domains). The following section provides... Mathematical guarantee of prediction error in the model: its upper limit of error is determined by... Modeling quality and and The divergence between them is jointly determined.

[0068] Definition 1. According to and Defined probability error The results are given by formulas (11) and (12) respectively:

[0069] in express j The actual power output after each time step.

[0070] Definition 2. Let This is a hypothetical distribution. Assumption According to Another distribution P The probability of inconsistency is defined by formula (13):

[0071] in Represents the XOR function. .

[0072] Definition 3. Given a domain and Distribution on and ,set up for The hypothetical class above. and The divergence can be defined as (14):

[0073] Definition 4. and Ideal joint hypothesis The assumption that minimizes the combination error is (15):

[0074] Lemma 1. For a field Any assumption on and ,

[0075] Proof: According to the probability inequality, when hour,

[0076] Therefore, And the proof of formula (16) is complete. Meanwhile, formula (17) can be proven using the following inequality:

[0077] Proposition 1: , in It is a constant.

[0078] Proof: According to Lemma 1 and Definition 3,

[0079] Therefore, It is an ideal assumption The overall error. In order to minimize It needs to be done through a probability estimator minimize and through feature extractor minimize (Its generator domain is invariant). Meanwhile, to ensure that Proposition 1 always holds, the discriminator... Generated hypothesis class It needs to be sufficiently rich and satisfy .

[0080] set up Then we have:

[0081] Therefore, when the discriminator With sufficient complexity, Judged as 0 When the judgment is 1), the following can be obtained. The upper bound of the condition. At this point, better performance can be obtained as shown in Proposition 1.

[0082] The DIFL framework constructs a network-based prediction model system distributed across client and server sides, executing iteratively through training. The specific iterative process is detailed below. Each iteration includes... Figure 2 The four steps shown are steps 1-4.

[0083] Step 1 (Forward Propagation): Based on the first... i One renewable energy power plant (client) iLocal data, processed by a local feature extractor Formula (8) is used to obtain local latent features. Subsequently, a local probability estimator was used. decoding Generate probability prediction (As described in formulas (6) and (7)). Discriminator on the server. By classifying the domain labels of local features, the local feature extractor is assisted in identifying domain-invariant features.

[0084] Step 2 (Model Aggregation): Global Feature Extractor on the Server With probability estimator Using formulas (9) and (10) respectively and By averaging the values, knowledge aggregation of the local model can be achieved.

[0085] Step 3 (backpropagation): Based on the combined loss of probability prediction and prediction domain label (Formula (28)), update the network parameters serving the local feature extractor and the local probability estimator through the backpropagation algorithm.

[0086] Step 4 (Model Output): Local Feature Extractor and probability estimator Use the server and value.

[0087] Through the four steps described above, the client can transfer knowledge without compromising data privacy. The following section will explain the structure of the client and server.

[0088] like Figure 2 As shown on the left, N Each independent client is responsible for handling private data. And provide domain-invariant features With probability prediction Each client consists of two collaborative networks: a local feature extractor and a local feature extractor. With local probability estimator .

[0089] In each training iteration m In the middle, local feature extractor Private inputs are obtained through formula (18). Convert to local features .

[0090]

[0091] In practical applications, to extract latent features from local data, four widely used deep networks are considered as... Candidate solutions include: fully connected neural networks (DNN), convolutional neural networks (CNN), long short-term memory networks (LSTM), and circular update unit networks (GRU).

[0092] Furthermore, local probability estimators For a mixed density network, it is as follows: Figure 3 The fully connected neural network shown outputs component parameters and local feature-based parameters. Provide prediction results Specifically, it is described by formulas (19) and (20).

[0093]

[0094] For each of In other words, It satisfies formula (3), and For parameters The distribution of the beta distribution. In practical applications, mixture models typically use the beta distribution as a component.

[0095] like Figure 2 As shown on the right, the server consists of three collaborative networks: a global feature extractor. Global probability estimator and discriminator .

[0096] In the m During each learning iteration and Perform the following steps using formulas (21) and (22) respectively. and The collection and aggregation of parameters.

[0097]

[0098] Meanwhile, the discriminator Try to distinguish local features using formula (23) The domain tag.

[0099]

[0100] In the m At the end of the next iteration, the server will and Distributed to multiple clients, these parameters will serve as the first... m The local model of +1 iteration is shown in equations (24) and (25).

[0101]

[0102] like Figure 4 As shown, the DIFL method uses... I The optimal parameters for the local and global models are obtained in the second training iteration. m In the nth iteration, the 1st i Local probability estimator for each client Using local training data , trained using formula (26) to maximize probability prediction The log-likelihood function.

[0103]

[0104] In addition, the discriminator By using feature extractors from different domains through formula (27) Generated features Train to maximize predicted labels equals field label The log-likelihood function.

[0105]

[0106] Meanwhile, local feature extractor Training is performed using formula (28), combined with the combined loss of formulas (26) and (27).

[0107]

[0108] Finally, based on the validation set performance defined by formula (29), from I A model and Select the optimal local model respectively .

[0109]

[0110]

[0111] After training the local feature extractor and probability estimator, the DIFL framework can use the constructed model system on the test set. ( The test was conducted on the device, and the prediction results were obtained using formulas (30) to (32). .

[0112]

[0113] This method was validated on two datasets: WFD (Wind Dynamics Dataset) and SPD (Saturation Dataset). The former was collected from six commercial wind farms in mainland China, and the latter from six grid-connected solar power plants. Both datasets contain historical power generation measurements and numerical weather prediction data (including temperature, air pressure, heat flux, radiation flux, precipitation, wind speed, and wind direction) from January 2019 to December 2020, with a sampling interval of 10 minutes. The processed data was divided into training, validation, and test sets, with proportions of 80%, 10%, and 10%, respectively.

[0114] The DIFL framework is implemented in PyTorch and GPU acceleration is enabled. Training is performed on a single NVIDIA GTX 2080Ti GPU. Performance evaluation adopts a widely accepted metric system, including Predicted Interval Coverage Probability (PICP), Predicted Interval Normalized Mean Width (PINAW), Average Coverage Error (ACE), Continuous Ranking Probability Score (CRPS), Normalized Root Mean Square Error (NRMSE), and Normalized Root Mean Square Error (NMAE), the specific expressions of which are shown in formulas (33)–(38):

[0115] in Indicates the number of samples. This represents the prediction interval at a specific confidence level. This represents the estimated cumulative distribution function. This represents the mode of the prediction. This represents the normalized actual wind power output, and PINC represents the nominal confidence level of the prediction interval.

[0116] To determine high-quality hyperparameter settings for training different models, a common grid search process was employed. The candidate hyperparameter settings considered are detailed in Table 2. These settings were derived by comprehensively considering preliminary experimental results and empirical knowledge from deep learning-based renewable energy prediction research. The performance of the two algorithms was evaluated based on the CRPS metric using a validation dataset.

[0117] Table 2: Hyperparameters and their setting options

[0118] To verify the method provided in the exemplary embodiments of the present invention, benchmark tests are constructed using the following different learning paradigms: RPF trained on local data (RPF-local): This method adopts the aforementioned general RPF learning paradigm and trains using only local data. RPF trained on full data (RPF-full): The data collected by the client is directly merged into a global dataset, and a general RPF learning paradigm is trained using this global dataset; FedAvg

[36] : The FedAvg framework is developed iteratively based on the horizontal joint learning paradigm. In each iteration, local models are trained using local data collected by the client. The global model on the server side aggregates the parameters of these local models by averaging. At the start of the search iteration, the parameters of the global model are copied to each local model.

[0119] The exemplary embodiments of the present invention first verify the performance of the DIFL paradigm and the benchmark model using CRPS, NMAE, and NRMSE metrics, as shown in Tables 3 and 4. As can be seen from the data on the left side of Table 3, all models trained using DIFL achieve the lowest mean CRPS metric based on WFD. Furthermore, by basing the results on each… The CRPS improvement percentage of DIFL relative to RPF-local, RPF-full, and FedAvg models was calculated separately for each candidate, and then the four models were compared. The average of the candidate parameters shows that DIFL achieves an average CRPS improvement of 5.49%, 10.55%, and 2.72% compared to RPF-local, RPF-full, and FedAvg, respectively. This is further demonstrated when using LSTM, GRU, and CNN to develop feature extractors. In this case, FedAvg becomes the second-best solution, outperforming the other three solutions. If DNN is used as... Therefore, RPF-local becomes a suboptimal solution. It should also be noted that, when based on WFD, different feature extractors... Similar results were obtained. It is worth emphasizing that a DNN was used as the feature extractor. At this point, DIFL's minimum performance still outperforms the best performance of RPF-local and RPF-full, and slightly outperforms the best performance of FedAvg, indicating that the learning paradigm selection is more critical than the feature extractor selection. LSTM is used as the feature extractor. The best performance can be obtained by combining it with DIFL as the learning paradigm. It is worth noting that RPF-full performs poorly in most scenarios, indicating that directly using datasets from all different sources does not guarantee better performance, further validating the importance of domain-invariant features extracted by the DIFL paradigm. Similarly, the CRPS values ​​of different SPD-based methods are shown on the right side of Table 3. It can be observed that DIFL outperforms other baseline methods, achieving improvements of 8.17%, 22.36%, and 4.02% compared to RPF-local, RPF-full, and FedAvg, respectively. It is also noted that the GRU model outperforms other models in all three paradigms: RPF-local, FedAvg, and DIFL. GRU is used as... The best performance can be obtained by combining it with DIFL as the learning paradigm. It is worth noting that the RPF-full paradigm performs the worst, indicating that simply merging local datasets may increase noise and impair the quality of predictive model development.

[0120] Table III: CRPS values ​​for different methods

[0121] Table 4: NMAE values ​​for different methods

[0122] Table 5: NRMSE values ​​for different methods

[0123] As shown in Table 4, DIFL also significantly outperforms the baseline model in terms of NMAE metrics. Compared to RPF-local, RPF-full, and FedAvg, DIFL achieves improvements of 4.7%, 10.2%, and 4.2% respectively based on WPD, and 5.1%, 30.2%, and 18.0% respectively based on SPD. Table 4 also shows that DIFL achieves the best NRMSE values ​​for both WPD and SPD. Based on WPD, DIFL outperforms RPF-local, RPF-full, and FedAvg by 6.3%, 11.1%, and 5.6% respectively. Based on SPD, DIFL achieves improvements of 6.0%, 30.0%, and 18.2% compared to the baseline scheme.

[0124] As can be seen from Tables 3 and 4, DIFL performs best in most scenarios across all three metrics. This observation suggests that despite the heterogeneous characteristics of data patterns from different types of renewable energy or renewable power plants, DIFL's flexibility enables it to address this data-driven modeling challenge. However, simply integrating data from multiple renewable energy power plants into a large training set may result in lower modeling performance than modeling schemes that only use local data. Models trained with integrated data may overfit to mixed information rather than the information features of a specific system. By introducing domain labels and a discriminator that distinguishes domain labels, DIFL achieves domain-invariant feature learning while ensuring privacy. This innovation may be a key factor in optimizing DIFL's predictive performance. Furthermore, this study also compares the performance of DIFL with the current state-of-the-art model—DLinear

[51] trained on local data (results are shown in Table 7). As can be seen from Table 7, DIFL outperforms DLinear in most cases. Compared to DLinear, DIFL shows an average improvement of 2.56%, 4.74%, and 6.60% in WFD-based CRPS, NMAE, and NRMSE metrics, respectively, and an improvement of 0.52%, 3.36%, and 6.62% in SPD-based metrics.

[0125] Table 6: Advances of DIFL compared to different methods

[0126] Table 7: Comparison between DIFL and DLinear

[0127] Furthermore, to explore the importance of choosing different modeling paradigms or model structures for developing effective DRPSF models, a comparative data analysis was conducted, and the results are shown in Table 8. Two analysis settings were designed: DAS1 and DAS2. In DAS1, the CRPS, NMAE, and NRMSE test results for all model structures were averaged according to the four modeling paradigms, thus obtaining the average CRPS, average NMAE, and average NRMSE for each modeling paradigm. In DAS2, the CRPS, NMAE, and NRMSE test results based on all modeling paradigms were averaged according to the four model structures. This yielded the average CRPS, average NMAE, and average NRMSE values ​​for DAS2. Under this design, DAS1 represents the analysis focusing on the modeling paradigm, while DAS2 represents the analysis focusing on the selected model structure. The standard deviations (St. dev.) and ranges of the obtained average CRPS, average NMAE, and average NRMSE can be further calculated. As shown in Table 8, the standard deviations and ranges based on the modeling paradigm selection are significantly higher than those based on the model structure selection. This finding indicates that, compared to model structure selection, prudent modeling paradigm selection has a more significant impact on improving DRPSF prediction performance. Furthermore, as shown in Tables 3 to 6, in most cases, the DIFL paradigm achieves comparable or better performance based on the considered model structure, further validating the value and effectiveness of DIFL.

[0128] Table 8: The Importance of Modeling Paradigm Selection and Model Structure Selection in DRPSF

[0129] Table 9: Performance of constructing metrics when PINC=0.9

[0130] To further evaluate performance, we then presented the prediction interval (PIC) and prediction interval error (PINAW) generated by estimating the probability density function (PDF) using different modeling paradigms, with PINC set to 0.9. The results are shown in Table 9. It can be observed that in most scenarios, the DIFL method achieves the best or second-best PICP, while the PINAW is at a moderate level.

[0131] Table 10: Performance of constructing prediction intervals when PINC=0.9

[0132] To visualize the PI values ​​obtained by DIFL, PI curves constructed using RPF-local, DIFL, and FedAvg were plotted based on the WFD and SPD datasets, respectively, using LSTM as the feature extractor (see [link to dataset]). Figure 5 and Figure 6It can be observed that the PI constructed by DIFL fits the power generation sequence better. Meanwhile, in Figure 7 In this study, the relationship curves between PICP and PINAW were plotted based on randomly selected power plants in the learning paradigm test set. It can be observed that when PINAW approaches the upper limit of wind speed normalization, PICP tends to 1; while DIFL can obtain a similar PICP at a smaller PINAW value.

[0133] Finally, the robustness of DIFL was tested through performance evaluations in different seasons. Based on the aforementioned calculation results, LSTM is considered... The ideal candidate model is DIFL. Four modeling paradigms were applied and compared: RPF-local, RPF-full, FedAvg, and DIFL, with results shown in Table 8. As shown in Table 10, DIFL still performs best in most scenarios. Therefore, compared to the other three benchmark models, DIFL is less sensitive to seasonal changes. Based on the above experimental results, it can be concluded that DIFL achieves state-of-the-art performance in the RPF task studied above.

[0134] In summary, the exemplary embodiments of this invention propose an innovative framework based on domain-invariant feature learning (DIFL) to address a challenging yet advanced task—the next-day probability prediction of renewable energy sequences while maintaining privacy. The DIFL method enables knowledge sharing across multiple sites while avoiding the leakage of local data at each site. DIFL consists of multiple clients and one server. Each client develops a feature extractor to encode the input into latent features and designs a probability estimator to provide probability predictions based on the extracted latent features. On the other hand, the server has a dual function: 1) aggregating the knowledge of the local model and distributing the aggregated parameters back to the clients; 2) distinguishing the domain labels of features through a discriminator, assisting the feature extractor in generating domain-invariant features.

[0135] To validate the advantages of DIFL, an upper bound for the prediction error was derived through mathematical analysis. This upper bound is jointly constituted by the source domain modeling quality and the difference between the source and target domains. Data from six commercial wind farms and six solar power plants were used in the experiments. Comparison with several well-known methods confirms that DIFL has reached the state-of-the-art level. However, performance evaluation under different meteorological conditions was not addressed here because the dataset used did not include meteorological information. Future research could collect meteorological data to further explore this important topic. Furthermore, the current version of DIFL uses only a single network architecture design to model all renewable energy power plants, although its predictive performance has improved. Future efforts should focus on extending DIFL to support different network architecture designs to model different types of power plants while protecting privacy. Particular attention will be paid to researching novel modeling principles that can efficiently select appropriate model structures, thereby improving modeling flexibility to capture the heterogeneity of individual power plants.

[0136] Various method embodiments of the present invention can be implemented using systems implemented in hardware and / or software. For example, Figure 8 A data processing system 300 is illustrated in some embodiments of the present invention. This system can be used to perform the aforementioned method for predicting the previous day's probability of renewable energy sequences. More broadly, the data processing system 300 can be used to perform or assist in performing one or more method embodiments of the present invention.

[0137] Data processing system 300 typically includes necessary components to receive, store, and execute appropriate computer instructions, data, commands, and / or code. The main components of data processing system 300 include processor 302 and memory (storage unit) 304. Processor 302 may include one or more of the following: CPU, MCU, GPU, logic circuit, Raspberry Pi chip, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or any other digital or analog circuit configured to interpret and / or execute program instructions and / or process signals and / or information and / or data. Memory 304 may include one or more volatile memories (such as RAM, DRAM, SRAM, etc.), one or more non-volatile memories (such as ROM, PROM, EPROM, EEPROM, FRAM, MRAM, FLASH, SSD, NAND, NVDIMM, etc.), or any combination thereof. Suitable computer instructions, commands, codes, information, and / or data may be stored in memory 304. Computer instructions used to perform or assist in performing embodiments of the methods of the present invention may be stored in memory 304. The processor 302 and the memory (storage device) 304 can be integrated or separated (and linked together through an operable connection).

[0138] Optionally, the data processing system 300 may also include one or more input devices 306. Examples of such input devices 306 include: keyboards, mice, styluses, image scanners, microphones, haptic / touch input devices (such as touchscreens), image / video input devices (such as cameras), etc. Input devices 306 may be used to receive user input. Optionally, the data processing system 300 may also include one or more output devices 308. Examples of such output devices 308 include: displays (such as screens, projectors, etc.), speakers, headphones, printers, additive manufacturing equipment (such as 3D printers), etc. Displays may include liquid crystal displays, LED / OLED displays, or other suitable display devices, which may or may not have touch sensing capabilities. Output devices 308 (such as displays) may be used to present three-dimensional medical images, raw slide images, reconstructed slide images, residual slide images, etc. The data processing system 300 may also include one or more disk drives 312, which may include one or more of the following: solid-state drives, hard disk drives, optical drives, flash drives, tape drives, etc. The data processing system 300 can be equipped with a suitable operating system, such as one deployed on disk drive 312 or memory 304. Memory 304 and disk drive 312 are jointly controlled by processor 302. The data processing system 300 may optionally be equipped with communication device 310 for establishing one or more communication links (not shown) with other computing devices (such as servers, personal computers, terminals, tablets, mobile phones, smartwatches, IoT devices, or other wireless computing devices). Communication device 310 may include one or more of the following components: modem, network interface card (NIC), integrated network interface, NFC transceiver, ZigBee transceiver, Wi-Fi transceiver, Bluetooth® transceiver, RF transceiver, cellular (2G, 3G, 4G, 5G, 5G and above, etc.) transceiver, optical port, infrared port, USB connection, or other wired / wireless communication interface. The transceiver may be implemented by one or more devices (integrated transceiver, standalone transceiver, etc.). The communication link may transmit commands, instructions, information, and / or data via wired or wireless means. For example, processor 302 and memory 304 (optionally including input device 306, output device 308, communication device 310, and disk drive 312 (if present)) may be directly or indirectly interconnected via a bus, peripheral component interconnect (PCI) such as PCI Express, Universal Serial Bus (USB), Fibre Channel bus, or other means. In one embodiment, at least some components may be wirelessly connected, for example, via a network such as the Internet or a cloud computing network.

[0139] Those skilled in the art should understand that, Figure 8 The data processing system 300 shown is merely an example; in other embodiments, the system may employ different configurations (e.g., including additional components, reducing the number of components, etc.).

[0140] While not strictly necessary, one or more embodiments described with reference to the accompanying drawings can be used by developers as an application programming interface (API) or a set of libraries, and can also be integrated into other software applications, such as terminal or computer operating systems, portable computing device operating systems. In one or more embodiments, when a program module comprises routines, programs, objects, components, and data files, those skilled in the art will understand that the functionality of a software application can be distributed across multiple routines, objects, and / or components to achieve the functionality required by the present invention.

[0141] The exemplary embodiments of the present invention have been fully described above. Although the specification refers to specific embodiments, those skilled in the art will understand that the present invention can be modified in terms of specific details without changing the core technical solution. Therefore, the present invention should not be construed as being limited to the embodiments described herein.

[0142] While the embodiments have been described in detail in the accompanying drawings and foregoing description, they should be understood to be illustrative rather than restrictive, and the examples shown are merely exemplary embodiments and do not limit the scope of the invention in any way. It should be understood that any feature described herein can be applied to any embodiment. Exemplary embodiments are not mutually exclusive, nor do they exclude other embodiments not listed herein. Accordingly, the present invention also provides embodiments comprising one or more combinations of the illustrative embodiments described above. Modifications and variations to the content of the present invention are acceptable as long as they do not depart from its spirit and scope. Therefore, limitations should only be imposed based on the content shown in the appended claims.

Claims

1. A computer-implemented method for probabilistically predicting the next-day power generation sequence of multiple renewable energy power plants while protecting data privacy, the method comprising: a) For each client device, map its raw data input to potential features; The plurality of client devices correspond to each of the plurality of renewable energy power plants; b) Transmit the locally hosted prediction model of each client device to the server in the form of the latent features of the client device and the model parameters; all of the client devices are connected to the server; c) Aggregate the locally hosted prediction models of the multiple client devices at the server; d) Distribute the aggregated model to the multiple client devices; e) Based on the aggregation model, update the localized prediction model on each of the plurality of client devices; and f) On each of the plurality of client devices, generate a probabilistic prediction of the power output sequence based on the updated localized prediction model.

2. The computer-implemented method of claim 1, wherein for each of the plurality of client devices, step a) is performed by a local feature extractor on that client device.

3. The computer-implemented method of claim 2, wherein the local feature extractor is a deep neural network, a convolutional neural network, a long short-term memory network, or a gated recurrent unit.

4. The computer-implemented method of claim 2, wherein in step a), the local feature extractor is assisted by a discriminator on the server to identify domain-invariant features.

5. The computer-implemented method of claim 1, wherein the latent feature is a domain-invariant feature.

6. The computer-implemented method of claim 1, wherein the model parameters are generated by a local probability estimator of each of the plurality of client devices.

7. The computer-implemented method of claim 1, wherein the server comprises a global feature extractor, a global probability estimator, and a discriminator.

8. The computer-implemented method of claim 7, wherein the global feature extractor is adapted to aggregate all the latent features from the plurality of client devices; and the global probability estimator is adapted to aggregate all the model parameters from the plurality of client devices.

9. The computer-implemented method of claim 7, wherein the aggregation model includes aggregated latent features and aggregated model parameters for updating local feature extractors and local probability estimators on the plurality of client devices.

10. The computer-implemented method of claim 7, wherein the discriminator is adapted to classify the domain labels of the latent features.

11. The computer-implemented method of claim 1, wherein steps a)-e) are repeatedly executed through multiple iterations to train the locally hosted prediction model.

12. The computer-implemented method of claim 6, further comprising the following steps: g) Using a training dataset and a validation dataset, train the local probability estimator on at least one of the client devices to maximize the log-likelihood of the probability predictions made by the local probability estimator.

13. The computer-implemented method of claim 7, further comprising the following steps: h) The discriminator is trained using features generated by multiple local feature extractors located on the multiple client devices in different domains to maximize the log-likelihood of the predicted label being equal to the domain label.

14. The computer-implemented method of claim 2, further comprising: The local feature extractor on at least one of the client devices is trained using a combined loss derived from training the local probability estimator on at least one of the client devices and training the discriminator on the server.

15. A system for probabilistically predicting the next-day power generation sequence of multiple renewable energy power plants, the system comprising: One or more processors; as well as A memory storing one or more programs executed by the one or more processors, the one or more programs containing instructions adapted to perform the method according to claim 1.

16. A system for probabilistically predicting the next-day power generation sequence of multiple renewable energy power plants, the system comprising: server; as well as Multiple client devices connected to the server; The server is adapted to aggregate locally hosted prediction models from the multiple client devices and distribute the aggregated models to the multiple client devices. The locally hosted prediction models received by the server include the potential features and model parameters of these locally hosted prediction models.

17. A non-transitory computer-readable storage medium storing one or more programs suitable for execution by one or more processors, the one or more programs containing instructions suitable for performing the method of claim 1.

Images