Systems and methods for energy production asset performance analysis

Abstract

A system and method for image-based energy asset performance analytics are described. The system transforms time-series energy production data of an energy asset into a first image representing an amount of energy produced by the energy asset as a function of time and generates a first residual image using a first autoencoder model trained to remove effects associated with a first asset performance factor from the first image. A determination of a contribution of the first asset performance factor to a performance loss of the energy asset over time can be made based on a comparison of the first residual image to the first image.

Description

SYSTEMS AND METHODS FOR ENERGY PRODUCTION ASSET PERFORMANCE ANALYSIS CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S Provisional Patent App. No.63 / 750,225, filed January 27, 2025, which is hereby incorporated by reference herein. TECHNICAL FIELD

[0002] This invention relates generally to energy production assets and, more specifically, to solar energy production asset performance analytics.BACKGROUND

[0003] Energy production asset performance can vary significantly depending on many internal and external factors, such as hardware configuration, the condition of the parts, and weather conditions. A clear insight into the factors affecting the performance of energy assets is important to troubleshooting, maintenance, and repair of existing energy assets, as well as decisions around future asset deployments.BRIEF DESCRIPTION OF DRAWINGS

[0004] Various needs are at least partially met through provision of the system and methods for energy production asset performance analysis described in the following detailed description, particularly when studied in conjunction with the drawings. A full and enabling disclosure of the aspects of the present description, including the best mode thereof, directed to one of ordinary skill in the art, is set forth in the specification, which refers to the appended figures, in which:

[0005] FIG. 1 is a block diagram of an image-based asset analytics system in accordance with some embodiments.

[0006] FIG. 2 is a block diagram of a data pipeline in accordance with some embodiments.

[0007] FIG. 3 is a graph showing energy production over time in accordance with some embodiments.Attorney Docket No. 21830-160619-US

[0008] FIG. 4 is a graph of energy production data in accordance with some embodiments.

[0009] FIG. 5 is the image of FIG. 3 with weather effects removed in accordance with some embodiments.

[0010] FIG. 6 is the image of FIG. 3 with shade effects removed in accordance with some embodiments.

[0011] FIG. 7 is the image of FIG. 3 with hardware failure effects removed in accordance with some embodiments.

[0012] FIG. 8 is a graph showing a comparison of energy production levels with effects removed in accordance with some embodiments.

[0013] FIG. 9 includes graphs showing a comparison of energy production levels over a period of time in accordance with some embodiments.

[0014] FIG. 10 is a block diagram of an exemplary computing system in accordance with some embodiments.

[0015] FIG. 11 is a flow diagram of an exemplary method in accordance with some embodiments.

[0016] Elements in the figures are illustrated for simplicity and clarity and have not necessarily been drawn to scale. For example, the dimensions and / or relative positioning of some of the elements in the figures may be exaggerated relative to other elements to help to improve understanding of various embodiments of the present teachings. Also, common but well-understood elements that are useful or necessary in a commercially feasible embodiment are often not depicted in order to facilitate a less obstructed view of these various embodiments of the present teachings. Certain actions and / or steps may be described or depicted in a particular order of occurrence while those skilled in the art will understand that such specificity with respect to sequence is not actually required.Attorney Docket No. 21830-160619-USDETAILED DESCRIPTION

[0017] Generally speaking, pursuant to various embodiments, systems, apparatuses, and methods are provided herein for analyzing energy asset performance. In some embodiments, an image-based energy asset performance analytics system is configured to determine a plurality of energy asset performance parameters based on energy output data of an energy asset.

[0018] The following description is not to be taken in a limiting sense, but is made for the purpose of describing the general principles of exemplary embodiments. Reference throughout this specification to "one embodiment," "an embodiment," "some embodiments," "an implementation," "some implementations," "some applications," or similar language means that a particular feature, structure, or characteristic described in connection with an embodiment is included in at least one embodiment of the invention. Thus, occurrences of these phrases throughout the specification may, but do not necessarily, all refer to the same embodiment.

[0019] In this description, technical terms are used consistently with their ordinary meanings in the field, except where explicitly defined otherwise. As used herein, the word "or" is intended to be inclusive, meaning "and / or," unless the context clearly dictates otherwise. Thus, a condition of "A or B" is satisfied by A alone, B alone, or both A and B. The singular forms "a," "an," and "the" include plural references unless the context clearly dictates otherwise. Additionally, terms like "coupled" or "attached" (or "fixed") encompass both direct and indirect connections (through intermediate elements) unless otherwise specified. Approximating language such as "about" or "approximately," when used to modify a quantitative value, means that minor variations (e.g., within 10%) of the stated value are within the intended scope.

[0020] Performance and efficiency of energy production assets (referred to simply as "energy assets" herein) can be affected by many factors. Current methods for energy asset monitoring, particularly for solar energy assets, have various inefficiencies.Common approaches include using a hypothetical asset model (e.g., a digital twin) toAttorney Docket No. 21830-160619-UScompare expected vs. actual energy production; using self-reported loss codes from hardware (e.g., OEM inverter codes); and using signal decomposition techniques like Signal Level Attenuation Characterization (SLAC). Digital twin methods generally require extensive data inputs and detailed hardware information that are often inaccessible, incomplete, or inaccurate. OEM codes frequently produce false positives or false negatives and are often unreliable and inefficient. SLAC implementations typically require additional fine-tuning, have inefficient inference times, and fail to account for all potential performance loss factors. Accordingly, there is a need for an improved method of energy asset performance monitoring.

[0021] The present disclosure provides an image-based analytics system and method that uses machine learning models (MLMs) to monitor the performance of an energy asset. Compared to previous methods, the disclosed system can derive performance analytics based only on the energy output signal from the asset and does not require detailed hardware specifications. The MLMs can be trained to recognize and manage data quality issues and to attribute a root cause to performance shortfalls. Further, embodiments of the system require no manual parameter tuning, have a quick and efficient inference time (e.g., on a standard CPU the system can run in ~2 seconds, and on a GPU in ~300 milliseconds), and account for performance losses due to root causes including weather, shade, soiling, and hardware issues. Many existing analytics solutions that evaluate performance and losses from energy production data approach the analysis as a time-series problem. In contrast, the disclosed system uses an imageprocessing approach, employing deep learning algorithms such as autoencoders to evaluate energy asset performance.

[0022] MLMs in this system output features that allow the analytics to automatically determine the root cause of observed symptoms (performance issues) on an energy asset. An automated root-cause identification system enables issues affecting asset performance to be identified and prioritized for mitigation, thereby streamlining maintenance task assignment. The energy asset performance analytics system providesAttorney Docket No. 21830-160619-USa fault-tolerant pipeline that sequentially ingests and transforms asset data, detects and diagnoses issues, organizes agent work, and hosts remediation workflows. The system can utilize a secondary model chain that leverages historical weather data, deterministic physics-based models, and simulated losses to generate a synthetic dataset (for model training and calibration). Overall, the image-based analytics approach can provide reduced processing speed, improved precision and recall for the root-cause categories addressed (weather, shading, soiling, hardware). Further, the system can provide organizations that own or operate energy assets with a portfolio-level view of energy production data, including estimated losses attributable to each root cause for their solar assets.

[0023] To support different deployment scenarios, the image-based energy asset performance analytics system can be implemented in various forms. For example, in some embodiments, the functionality is implemented as a software application running on a server or cloud platform that remotely monitors solar installations. In other embodiments, it may be built into a local controller at the asset site. In either case, because the system requires only the energy output data from the asset (and no detailed hardware specifications), it can be deployed flexibly as a standalone remote service or integrated on-site, making the analytics available for a wide range of scenarios, from individual installations to large portfolios of assets. The instructions executed by the processor can reside on any suitable non-transitory machine-readable medium (such as flash memory, a hard drive, or a solid-state disk), ensuring the system's logic is persistently stored and not merely a transitory signal. It should also be appreciated that the invention encompasses corresponding methods of image-based energy asset performance analysis and computer-readable media storing instructions consistent with the system features described.

[0024] The foregoing and other benefits may become clearer upon a thorough review of the following detailed description.Attorney Docket No. 21830-160619-US

[0025] Referring now to the figures, FIG. 1 illustrates an image-based energy asset performance analytics system 100 in accordance with various embodiments. Generally, the system 100 may be employed to determine the contribution of an asset performance factor to a performance loss of an energy asset over time. As used herein, "asset performance factors" refer to conditions or influences that can impact the energy output of an asset, including but not limited to weather conditions (e.g., cloud cover), shading from nearby objects, hardware faults (e.g., inverter or panel failures), and soiling (e.g., dust or debris accumulation). These performance factors correspond to root causes of energy loss, and the system is configured to isolate and quantify the impact of each factor using machine learning models. By attributing performance degradation to specific root causes, the system enables targeted diagnostics and informed maintenance planning, as described more fully below.

[0026] In various embodiments, system 100 may be implemented using one or more processors operating in a distributed or centralized computing environment. For example, the components of system 100 may be deployed across multiple physical or virtual machines, such as cloud-based servers, edge computing nodes, or on-premises controllers located at the energy asset site. In some implementations, data acquisition may occur locally at the asset, while more computationally intensive analytics (e.g., autoencoder inference and performance factor attribution) are performed remotely in a cloud infrastructure. In some embodiments, one or more functionalities of system 100 may be performed via service calls and communications with external service providers, enabling integration with third-party platforms or analytics services. Alternatively, the entire system may be hosted on a single computing device, such as a server or industrial PC, depending on the scale and requirements of the deployment. Generally, the system may be used to monitor individual assets and / or manage large-scale portfolios across geographically distributed sites.

[0027] The system 100 receives energy asset production data 102 as input. The energy asset production data 102 includes time-series measurements representing the asset'sAttorney Docket No. 21830-160619-USenergy output recorded at intervals over a defined period (for example, hourly energy production measurements over days, months, or years). This energy production data may be provided directly by the asset (e.g., via on-site meters or sensors) or retrieved from a data repository. In the context of a solar energy asset, the asset may include components such as photovoltaic (PV) cells or panels along with sensors or inverters that collect and transmit performance data. The production data 102 generally provides the power or energy output level (e.g., kilowatts or kilowatt-hours) of the asset over time.

[0028] In some embodiments, the system 100 includes an analytics system 104 that employs multiple deep learning models (MLMs), which may include a weather MLM 108, a shade MLM 110, a hardware failure MLM 112, and a soiling MLM 113. These MLMs can be implemented as neural networks (for instance, autoencoders or other architectures) trained to process the asset's performance data. Autoencoders are one useful type of model: they encode input data into a latent representation and then decode it back, thereby learning essential features of the input. In this system, the weather MLM 108, shade MLM 110, hardware failure MLM 112, and soiling MLM 113 are models that take images as input (as explained below) and are each trained to isolate the effects of a particular performance factor.

[0029] The energy asset production data 102 is provided to an image converter 106. The image converter 106 transforms the time-series energy output data into a two- dimensional image representation. FIG. 3 shows an example graph 102A of a solar asset's energy production (in kW) over time, in this case hourly production over a year. FIG. 4 shows an example image 106A created by the image converter 106 from that energy data. In this example, image 106A visualizes the asset's actual energy production as a function of time of day (on one axis) and day of the year (on the other axis). The data may be normalized by the PV system's size (the peak power under Standard Test Conditions, STC) to create a consistent relative scale. In image 106A, intensity indicates the level of energy produced (for instance, darker areas toward the center indicatingAttorney Docket No. 21830-160619-UShigher production, and lighter areas at the edges indicate little or no production during night hours or outages).

[0030] The weather MLM 108 is a deep learning model (e.g., an autoencoder) configured to analyze an input image that represents the asset's energy production (such as image 106A) and to isolate the effects of weather. In one example, the weather MLM 108 takes image 106A (which includes all effects) and produces a new image 108A (see FIG. 5) that represents the energy production with weather effects removed. In other words, weather MLM 108 is trained to estimate what the asset's performance would have been with no weather-related reductions (e.g., under clear sky conditions throughout). The difference between the original image 106A and the weather- corrected image 108A corresponds to the performance loss due to weather (for example, energy not produced because of cloud cover). In some embodiments, the output of the weather model can be further processed or de-normalized to yield an estimated "clear-sky" energy production profile for the asset.

[0031] The shade MLM 110 is a deep learning model (e.g., autoencoder) that accepts the output of the weather MLM 108 (image 108A, which has weather effects removed) and removes the effect of shading from that data. The result is an image 110A (see FIG.6) representing the asset's normalized energy production with shade effects removed. In essence, the shade MLM 110 is trained to determine the asset's performance as if no shading (from buildings, trees, or other obstructions) occurred, given the weather- corrected performance as input. The difference between image 108A and image 110A can be interpreted as the energy lost due to shading. In some embodiments, the shade MLM 110 is also configured to output a quantitative estimate of energy (in kWh) lost to shade over the period.

[0032] The hardware failure MLM 112 is a deep learning model (e.g., autoencoder) configured to remove the effect of hardware-related issues from the performance data. It takes the output of the shade model (image 110A) and produces an image 112A (see FIG. 7) representing the normalized energy production with hardware failure effectsAttorney Docket No. 21830-160619-USremoved. In other words, the hardware MLM 112 is trained to project what the asset's performance would be if there were no hardware problems, given the performance with weather and shading already accounted for. The difference between image 110A and image 112A corresponds to performance loss due to hardware issues (such as inverter outages, panel failures, or circuitry problems). Generally, image 112A has the same format and scale as the original image 106A, since it represents the asset's production under ideal conditions (no weather, no shade, no hardware faults) aside from soiling.

[0033] The soiling MLM 113 is a further deep learning model (e.g., autoencoder) configured to remove the effects of soiling (dust, dirt, snow, or other debris on the panels) from the performance data. It takes the output of the hardware model (image 112A) and produces an image representing the normalized energy production with soiling effects removed. In other words, given the performance with weather, shade, and hardware issues stripped out, the soiling MLM 113 estimates what the performance would be if the panels were clean. The difference between image 112A and this soiling- corrected image corresponds to the energy loss due to soiling. In some embodiments, the soiling MLM 113 can output the estimated energy (in kWh) lost to soiling.

[0034] While the embodiments above specifically describe models for weather, shade, hardware, and soiling factors, it is understood that the system could include models for additional or alternative factors, or omit certain factors, as needed. The order of applying these factor-specific models could vary in some implementations, and additional models may be introduced to handle other influences on performance (for example, high-temperature derating, snow cover, curtailment events, etc.). The architecture is flexible such that any asset performance factor that measurably affects output could have a corresponding MLM in the chain. Additionally, while FIG. 1 shows the MLMs processing in series, using the output of the prior MLM, in some embodiments, the MLMs may process the energy asset product data in parallel orAttorney Docket No. 21830-160619-USindividually, such that each MLM outputs an image with the respective effect removed only.

[0035] The machine learning models described above (108, 110, 112, 113) can be initially trained and periodically retrained using various data sources and feedback to improve their accuracy. Training data may include combinations of real operational data and simulated (or augmented) data. In general, training and retraining are performed to continuously improve the models' results and can be done at any suitable interval (for example, as new data becomes available or when performance drifts).

[0036] In some embodiments, the machine learning models (MLMs) described herein, such as the weather MLM 108, shade MLM 110, hardware failure MLM 112, and soiling MLM 113, are implemented as autoencoder-based encoders trained using techniques analogous to image denoising. Specifically, each MLM is trained to receive an input image representing energy production data affected by a particular asset performance factor and to reconstruct a version of the image with that factor's influence removed, similar to how a denoising autoencoder learns to reconstruct a clean image from a noisy input. This training approach enables each model to learn latent representations that isolate and suppress the visual patterns associated with specific performance degradations (e.g., cloud cover, shading artifacts, or hardware anomalies), thereby allowing the system to generate residual images that reflect the asset's performance under idealized conditions.

[0037] In some embodiments, the MLMs are trained using an unsupervised learning approach on paired data that represents performance with and without a given effect. For example, the weather MLM 108 can be trained using pairs of images where one image includes normal operation with weather variability and the other is a "clear-sky" version (e.g., generated by a physics-based clear-sky model or taken from a day with no clouds). In one implementation, historical operational data for an asset (or multiple assets) is paired with corresponding clear-sky simulations, allowing the weather model to learn to reconstruct the clear-sky output from the weather-affected input. ThisAttorney Docket No. 21830-160619-UStechnique enables the model to automatically learn features that distinguish weather impacts without needing explicit labels for cloudy vs. sunny conditions on each sample. Likewise, the shade MLM 110 and hardware MLM 112 can be trained on data sets where shading events or hardware faults are either known from logs or synthetically introduced, so that the models learn to remove those effects. The models can be validated on real-world data, for instance, by verifying that their outputs make sense for periods with known ground truth (such as an independently recorded outage or a monitored cleaning event). Over time, as more data from various assets and conditions is collected, the models 108, 110, 112 (and 113) can be fine-tuned or retrained to maintain and improve accuracy.

[0038] The analytics system 104 further includes an output comparison component 114, which compares the outputs of the various MLMs to the original input or the clear-sky output image (with each effect removed) to quantify each factor's contribution to performance loss. In practice, the output comparison 114 may calculate differences or ratios between the original performance image (from converter 106) and each successive residual image (108A, 110A, 112A, etc.). By doing so, it determines the magnitude of performance impact for each factor over time or in aggregate. For example, by comparing image 106A vs. 108A, the system quantifies the energy lost due to weather; comparing image 108A vs. 110A quantifies loss due to shade; and so on. The output comparison component 114 can thus produce metrics or time-series signals indicating how much loss is attributed to each root cause at each point in time (or over the entire analysis period). Significant factors can then be flagged as issues requiring mitigation. Alternatively, in embodiments where the MLMs process the input image in parallel, the output of each MLM may be compared with the input image individually to quantify the energy lost due to each factor.

[0039] FIG. 8 provides an illustration (graph 114A) of how the model outputs can be compared to yield insight. In FIG. 8, line 801 represents the ratio of the asset's performance with vs. without shade (e.g., actual output divided by the output if noAttorney Docket No. 21830-160619-USshade, i.e., 106A vs. 110A). Line 802 represents the ratio of the asset's performance with vs. without hardware failures (actual output divided by output if no hardware issues, i.e., 106A vs. 112A). In this example, line 802 stays near 1.0, indicating that removing hardware issues had negligible effect, implying hardware was generally functioning (no significant hardware-related losses). Line 801, however, drops significantly below 1.0 at certain times, indicating that the absence of shading would have substantially increased output, hence shading is causing a performance drop during those periods. For reference, line 803 illustrates the output of a prior-art analysis method, which is much noisier and shows spurious spikes (false positives) in detecting issues. This underscores the advantage of the layered autoencoder approach in filtering out noise and isolating true performance issues.

[0040] It should be understood that FIG. 8 is just one example of visualizing model outputs. In practice, any of the model results or comparisons could be graphed or reported in various ways to highlight performance issues. For example, actual vs. clearsky output may be plotted to show overall performance, or show cumulative energy loss per factor. Additional root cause models (beyond shade and hardware) could similarly have their outputs compared. Generally, by comparing the MLM outputs to the original data, the system can determine which factor (or factors) are causing observed underperformance.

[0041] In some embodiments, the output comparison component 114 is also configured to determine a year-over-year degradation rate of the asset using the MLM outputs over a long term. Degradation here refers to the long-term decline in the asset's ability to produce energy (for example, solar panel degradation). Typically this is measured in percentage or energy loss per year. The system can estimate degradation by examining trends in the "ideal" performance output over multiple years. Specifically, after removing short-term factors (weather, shade, etc.), any consistent downward trend likely indicates intrinsic degradation of the asset.Attorney Docket No. 21830-160619-US

[0042] To calculate degradation, the component 114 may use a clear-sky model or baseline and compare how the asset's maximum possible output changes from year to year. It can generate clear-sky performance predictions for each year (using the models 108-113 to strip out year-specific losses), and then compare those. A steady drop in the clear-sky predicted output each year would point to degradation. For example, if in Year 1 the clear-sky model indicates the system could produce X kWh under ideal conditions, and in Year 2 it indicates slightly less, the difference is due to degradation (assuming similar conditions and no new external factors). This year-over-year analysis can be done for both actual output (line 901 in FIG. 9, which includes weather variability) and theoretical clear-sky output (line 902, which excludes external factors).

[0043] FIG. 9 illustrates this concept. Graph 140 shows year-over-year actual power output under real conditions (subject to weather, etc.), and graph 141 represents the year-over-year theoretical power output with no losses (what the asset could have produced under ideal conditions), as derived from the MLM outputs. Both actual and ideal outputs can be normalized for fair comparison (accounting for any system changes or just to remove seasonal biases). If line 901 decreases over the years, it could be due to a combination of factors like unusually poor weather trends or gradual system issues; if line 902 is also decreasing, that specifically indicates the asset's maximum capacity is declining, i.e., true degradation (such as panel aging). In FIG. 9, suppose line 901 drops due to a very cloudy year, but line 902 remains the same, that would mean no degradation, just weather variation. If line 902 shows a decline (perhaps 2-3% over a couple of years), that quantifies the degradation rate of the asset. Knowing the degradation rate helps in forecasting maintenance (e.g., when to replace panels) and in predictive modeling of asset performance over its lifetime.

[0044] Turning now to the operational flow, FIG. 2 illustrates a data pipeline 200 for how the system 100 processes incoming data and acts on it. The pipeline 200 can be executed by one or more processors (such as those in a cloud server or on-site controller, see FIG. 10) running the system's software. The pipeline begins with theAttorney Docket No. 21830-160619-USenergy asset itself, labeled 202 (e.g., a solar PV system), which provides data to a data acquisition component 116. The data acquisition component 116 is responsible for collecting raw signals from the asset's sensors or meters and converting them into usable digital data. This may involve reading inverter telemetry, performing analog-to- digital conversion, filtering noise, aligning data timestamps, and scaling or unit conversion. Essentially, component 116 ensures that the raw data from the asset (voltage, current, power readings, etc.) is prepared for analysis.

[0045] The data serving component 118 stores and manages incoming data. It functions as be a database or repository that archives historical energy production data and makes it available for retrieval. The data may be used for trend analysis, visualization, or feeding into models.

[0046] The issue detection component 120 monitors the incoming performance data in real-time (or in periodic batches) to detect anomalies or deviations that suggest a performance issue may be occurring. The issue detection component 120 may itself use an MLM or other algorithm (for example, an anomaly detection model) to detect an issue when the performance is outside expected bounds such as the output dropping significantly below the clear-sky expectation, or if production flatlines.

[0047] In some embodiments, the model hosting 126 is a component that stores the machine learning models (e.g., 108, 110, 112, 113), such as a cloud server, and makes them accessible to the rest of the system, possibly over a network. This allows the analytics system 104 or the issue diagnosis component 128 to retrieve and apply the latest version of the models. It also enables updating the models in one place, for example, uploading a retrained model to the host so that all assets or instances of the system start using the improved model. Model hosting 126 can thus serve multiple sites or units, ensuring consistency in analytics across a fleet of assets.

[0048] The data stored by component 118, the alerts or intermediate results from detection component 120, and the relevant models from hosting 126 are then provided to an issue diagnosis component 128. The issue diagnosis component 128 performs theAttorney Docket No. 21830-160619-USroot cause analysis using the techniques above. For example, the issue diagnosis component 128 converts the input data into an image, runs it through a plurality of MLMs (e.g., weather, shade, hardware, soiling), and compares outcomes to determine what factors are contributing to the performance issue flagged by detection component 120.

[0049] Using the analysis, the issue diagnosis component 128 may break down, for example, an observed energy shortfall into percentages caused by weather vs. shade, etc. For instance, diagnosis component 128 may determine that a certain dip in performance is 70% explained by an unexpected shading event (maybe a new obstruction or an overgrown tree) and 30% by heavier clouds than usual, with no indication of hardware faults. In another scenario, it may find that everything looks normal except the clear-sky output is trending down, pointing to degradation. By pinpointing the main cause(s), the system moves beyond just detecting the presence of an issue and is able to identify the cause of the issue.

[0050] After diagnosis, the pipeline may trigger appropriate actions to address the identified issues. An operations agent triage component 122 receives information about the diagnosed issues and their root causes from component 128. The triage component 122 prioritizes these issues and assigns them for resolution, much like a ticketing system. For a company managing many assets, triage component 122 would rank issues so that critical problems (e.g., a major hardware failure at a large site) are addressed before minor ones (e.g., slight soiling on a small installation). It considers factors like severity of energy loss, number of assets affected, and time sensitivity.

[0051] Once prioritized, the issues are passed to a remediation workflow 124. The remediation workflow 124 organizes and tracks the tasks needed to resolve each issue. This could integrate with maintenance management systems or work order systems. For example, if the issue is heavy soiling on an array of panels, workflow 124 may generate a work order for a cleaning crew or schedule a cleaning event. If the issue is an inverter fault, it may schedule a technician visit and ensure a replacement unit is available. TheAttorney Docket No. 21830-160619-USworkflow 124 essentially ensures that identified issues lead to concrete actions: tickets are opened, responsible teams are notified, and progress is tracked until resolution.

[0052] For instance, suppose the system identifies that a particular solar farm has a persistent morning shading issue and also one string of panels with a failed inverter. The triage component 122 would likely mark the inverter failure as urgent (since that string is producing zero output) and the shading as less urgent (if it's partial shading and only for a short time daily). The remediation workflow 124 would then, for the inverter issue, automatically create a service task for an electrical technician, including details of which inverter and likely cause, possibly even ordering the replacement part in advance. For the shading, the workflow may create a task to investigate the cause (maybe trees that need trimming) and schedule it for a later date.

[0053] In some embodiments, the remediation workflow 124 can interface with external tools, for example, sending an email or SMS alert to a maintenance team, or logging the issue in an external asset management system.

[0054] In addition to scheduling human-driven interventions, the pipeline 200 can optionally take automated corrective actions in some cases. That is, the system may directly command the asset or related systems to mitigate an issue if possible. For example, if the analysis detects an impending performance drop due to a forecasted weather event (like an approaching storm), and the site has a battery storage, the system may preemptively charge the battery or adjust settings to optimize performance. Or if it identifies an inverter is gradually failing (but not completely out yet), it may send a command to switch to a backup inverter or re-route power flow in a way that preserves generation. Any such direct interventions may done via machine- readable commands sent to the asset or its control devices (which could be an API call to the inverter, a SCADA command, etc.).

[0055] Moving on to the computing infrastructure, FIG. 10 shows an exemplary hardware configuration for implementing the analytics system (or parts of it). This could represent a cloud server, an on-premises industrial PC, or any suitable computingAttorney Docket No. 21830-160619-USdevice. System 1000 includes at least one processor 1002 and memory 1004 (which stores instructions 1006 that the processor executes). It also includes storage 1008 for persistent data (like databases, logs, historical data). These components communicate via an internal bus 1010. The system 1000 may have an interface 1012 (such as network interfaces for Internet or LAN connectivity) and I / O ports 1014 for connecting to external devices. A transceiver 1016 may be present if wireless communication (like cellular or radio) is needed, for example, for remote sites.

[0056] The processor 1002 could be any standard CPU or microcontroller capable of running the algorithms and managing I / O. Memory 1004 would include both the program (software implementing the pipeline, models, etc.) and runtime data. Storage 1008 may hold the large volumes of historical data or any model files and configuration, which do not need to be in fast memory at all times. When the instructions 1006 (the system software, including the model computations and logic described) are executed by processor 1002, the computing system 1000 carries out the operations of imagebased performance analysis (data gathering, image conversion, model inference, comparison, etc., as described above).

[0057] Referring now to FIG. 11, a flow diagram of an exemplary method 1100 for image-based energy asset performance analytics is illustrated in accordance with some embodiments. The method begins at block 1110, where time-series energy production data of an energy asset is transformed into a first image representing the amount of energy produced by the asset as a function of time. In practice, this transformation can be performed by an image conversion component (such as the image converter 106 described above), which encodes the asset's output over a given period into a two- dimensional visual form. The first image serves as a compact performance representation including all factors affecting the asset's output (e.g., it may resemble image 106A of FIG. 4, depicting energy production intensity across hours of the day and days of the year). By converting raw time-series data into such an image, the systemAttorney Docket No. 21830-160619-USenables downstream machine learning models to readily identify patterns and anomalies in the asset's performance data.

[0058] Next, at block 1120, the method includes generating a first residual image using a first autoencoder model trained to remove effects associated with a first asset performance factor from the first image. In this step, the first image (containing the asset's actual performance with all influences) is fed into a factor-specific machine learning model, specifically, an autoencoder configured to isolate and eliminate the impact of one particular performance factor. The first autoencoder model processes the first image and produces a modified image in which the influence of the first asset performance factor has been substantially removed. Stated differently, the output of block 1120 is an image depicting the asset's performance as if the first factor had not affected it. For example, if the first asset performance factor is weather (such as cloud cover or irradiance variability), the first autoencoder model (analogous to the weather model 108 discussed previously) will generate an image that represents the energy asset's production under ideal weather conditions (e.g., clear sky), thereby filtering out weather-related reductions. The resulting first residual image is effectively a "cleaned" performance image with respect to that factor, showing what the asset's output would have been absent the performance impact of the first factor.

[0059] At block 1130, the method involves determining a contribution of the first asset performance factor to a performance loss of the energy asset by comparing the first residual image to the original first image. By analyzing the differences between these two images (with and without the first factor's effects), the system quantifies the extent to which the first factor has caused the asset's output to deviate from its potential. In essence, the performance loss attributable to the first factor is measured as the discrepancy between the actual performance (embodied in the first image from block 1110) and the performance with that factor removed (embodied in the first residual image from block 1120). For instance, continuing the weather example above, the difference between the original performance image and the weather-corrected residualAttorney Docket No. 21830-160619-USimage directly represents the energy shortfall due to weather influences (e.g., the lost production caused by cloudy conditions over the period of analysis). This comparison may be done on a pixel-by-pixel or aggregate basis to compute the magnitude of lost energy, yielding a clear indicator of how significantly the first factor contributed to underperformance over time. The output of block 1130 can be a metric or data set identifying the first factor's impact (for example, a certain percentage or number of kilowatt-hours of loss attributable to that factor), which can be used for further diagnostics and reporting.

[0060] It should be appreciated that FIG. 11 illustrates the process for a single asset performance factor for clarity; in practice, the method 1100 can be iteratively extended to multiple factors to fully analyze an asset's performance. After obtaining the first residual image (with the first factor removed) at block 1120, the same general process may be repeated for a second asset performance factor: the first residual image can be fed into a second autoencoder model (trained to remove effects of a second factor) to produce a second residual image, and by comparing the second residual image to the first residual image, the contribution of the second factor to performance loss can be determined. Likewise, a third factor can be addressed by applying a third autoencoder to the second residual image, yielding a third residual image, and comparing it with the second residual image, and so on. Through this cascaded sequence of model applications (each analogous to blocks 1120 and 1130 for different factors), the system can isolate the effects of multiple performance factors such as weather, shading, hardware faults, and soiling in succession. Each autoencoder in the sequence is preferably trained (for example, via unsupervised learning on paired data with and without the respective factor's influence) to specifically remove one factor's impact, ensuring that the residual images accurately reflect performance with that factor eliminated. By the end of this iterative process, the method yields a set of residual images and corresponding comparisons that quantify the contribution of each considered factor to the asset's performance loss. This information enables the image-Attorney Docket No. 21830-160619-USbased analytics system to identify the root causes of any detected performance shortfall and supports the prioritization of corrective actions (e.g., scheduling maintenance to address a dominant loss factor). In summary, the flow diagram of FIG. 11 encapsulates a novel analytic approach wherein raw energy production data is transformed into an image, successively processed by factor-specific autoencoder models to strip away the effects of various asset performance factors, and compared at each stage to determine how much each factor is responsible for observed losses in energy production.

[0061] Embodiments of the described system have been tested and have demonstrated the ability to identify root causes of performance issues with at least 85% accuracy. This high level of accuracy, combined with the system's automated and fast analysis, provides an efficient and reliable issue identification capability for energy assets using only their energy production data. By quickly pinpointing issues, the system significantly speeds up the mitigation process, which in turn helps to improve or restore the energy production of assets over time. In summary, this image-based analytics approach allows asset operators to detect, diagnose, and address performance problems more effectively than was previously possible using conventional methods.

[0062] Further aspects of the disclosure are provided by the subject matter of the following clauses:

[0063] An image-based energy asset performance analytics system comprising: a processor; and a machine readable medium storing instructions that, when executed by the processor, cause the processor to: transform a time-series energy production data of an energy asset into a first image representing an amount of energy produced by the energy asset as a function of time; generate a first residual image using a first autoencoder model trained to remove effects associated with a first asset performance factor from the first image; and determine a contribution of the first asset performance factor to a performance loss of the energy asset over time based on comparing the first residual image to the first image.Attorney Docket No. 21830-160619-US

[0064] The image-based energy asset performance analytics system of any preceding clause, wherein the instructions further cause the processor to: determine contributions of a plurality of asset performance factors using a plurality of autoencoder models each trained to remove effects associated with a different one of the plurality of asset performance factors.

[0065] The image-based energy asset performance analytics system of any preceding clause, wherein the plurality of asset performance factors includes two or more of weather conditions, shading, hardware faults, and / or soiling.

[0066] The image-based energy asset performance analytics system of any preceding clause, wherein the instructions further cause the processor to: generate a second residual image using a second autoencoder model trained to remove effects associated with a second asset performance factor from the first residual image; and compare the second residual image to the first residual image to determine a contribution of the second asset performance factor to the performance loss of the energy asset over time.

[0067] The image-based energy asset performance analytics system of any preceding clause, wherein the instructions further cause the processor to: generate a third residual image using a third autoencoder model trained to remove effects associated with a third asset performance factor from the second residual image; and compare the third residual image to the second residual image to determine a contribution of the third asset performance factor to the performance loss of the energy asset over time.

[0068] The image-based energy asset performance analytics system of any preceding clause, wherein the first autoencoder model is trained with an unsupervised machine learning algorithm using pairs of images corresponding to energy production data with and without effects associated with the first asset performance factor.

[0069] The image-based energy asset performance analytics system of any preceding clause, wherein the instructions further cause the processor to validate the first autoencoder model using real-world energy production data from energy assets.Attorney Docket No. 21830-160619-US

[0070] The image-based energy asset performance analytics system of any preceding clause, wherein the instructions further cause the processor to: generate a clear-sky prediction of the energy asset based on contributions of one or more asset performance factors, including the first asset performance factor, on the performance loss of the energy asset over time; and determine a degradation rate of the energy asset based on tracking clear-sky predictions of the energy asset over time.

[0071] The image-based energy asset performance analytics system of any preceding clause, wherein the instructions further cause the processor to generate a maintenance and / or service task for the energy asset based on contributions of one or more asset performance factors.

[0072] The image-based energy asset performance analytics system of any preceding clause, wherein the instructions further cause the processor to generate machine- readable command to the energy asset or an associated device to modify an operation of the energy asset.

[0073] A method for image-based energy asset performance analytics, comprising: transforming time-series energy production data of an energy asset into a first image representing an amount of energy produced by the energy asset as a function of time; generating a first residual image using a first autoencoder model trained to remove effects associated with a first asset performance factor from the first image; and determining a contribution of the first asset performance factor to a performance loss of the energy asset over time based on comparing the first residual image to the first image.

[0074] The method of any preceding clause, further comprising: determining contributions of a plurality of asset performance factors using a plurality of autoencoder models, each trained to remove effects associated with a different one of the plurality of asset performance factors.Attorney Docket No. 21830-160619-US

[0075] The method of any preceding clause, wherein the plurality of asset performance factors includes two or more of weather conditions, shading, hardware faults, and / or soiling.

[0076] The method of any preceding clause, further comprising: generating a second residual image using a second autoencoder model trained to remove effects associated with a second asset performance factor from the first residual image; and comparing the second residual image to the first residual image to determine a contribution of the second asset performance factor to the performance loss of the energy asset over time.

[0077] The method of any preceding clause, further comprising: generating a third residual image using a third autoencoder model trained to remove effects associated with a third asset performance factor from the second residual image; and comparing the third residual image to the second residual image to determine a contribution of the third asset performance factor to the performance loss of the energy asset over time.

[0078] The method of any preceding clause, wherein the first autoencoder model is trained with an unsupervised machine learning algorithm using pairs of images corresponding to energy production data with and without effects associated with the first asset performance factor.

[0079] The method of any preceding clause, further comprising: validating the first autoencoder model using real-world energy production data from energy assets.

[0080] The method of any preceding clause, further comprising: generating a clear-sky prediction of the energy asset based on contributions of one or more asset performance factors, including the first asset performance factor, on the performance loss of the energy asset over time; and determining a degradation rate of the energy asset based on tracking clear-sky predictions of the energy asset over time.

[0081] The method of any preceding clause, further comprising: generating a maintenance and / or service task for the energy asset based on contributions of one or more asset performance factors.Attorney Docket No. 21830-160619-US

[0082] A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to: transform time-series energy production data of an energy asset into a first image representing an amount of energy produced by the energy asset as a function of time; generate a first residual image using a first autoencoder model trained to remove effects associated with a first asset performance factor from the first image; and determine a contribution of the first asset performance factor to a performance loss of the energy asset over time based on comparing the first residual image to the first image.

[0083] Those skilled in the art will appreciate that the embodiments described above can be modified or varied without departing from the scope of the invention. The specific thresholds and performance factors can be tuned for different types of installations. Additional statistical tests or machine learning models could be incorporated to enhance detection. The architecture can scale to fleets of thousands of assets, with parallel processing of data. All such modifications are considered within the ambit of these teachings, as the invention is defined by the following claims and their equivalents, rather than the specific examples given in this description.Attorney Docket No. 21830-160619-US

Claims

What is claimed is:

1. An image-based energy asset performance analytics system comprising:a processor; anda machine readable medium storing instructions that, when executed by the processor, cause the processor to:transform a time-series energy production data of an energy asset into a first image representing an amount of energy produced by the energy asset as a function of time;generate a first residual image using a first autoencoder model trained to remove effects associated with a first asset performance factor from the first image; and determine a contribution of the first asset performance factor to a performance loss of the energy asset over time based on comparing the first residual image to the first image.

2. The system of claim 1, wherein the instructions further cause the processor to: determine contributions of a plurality of asset performance factors using a plurality of autoencoder models each trained to remove effects associated with a different one of the plurality of asset performance factors.

3. The system of claim 2, wherein the plurality of asset performance factors includes two or more of weather conditions, shading, hardware faults, and / or soiling.

4. The system of claim 1, wherein the instructions further cause the processor to: generate a second residual image using a second autoencoder model trained to remove effects associated with a second asset performance factor from the first residual image; and compare the second residual image to the first residual image to determine a contribution of the second asset performance factor to the performance loss of the energy asset over time.Attorney Docket No. 21830-160619-US5. The system of claim 4, wherein the instructions further cause the processor to: generate a third residual image using a third autoencoder model trained to remove effects associated with a third asset performance factor from the second residual image; and compare the third residual image to the second residual image to determine a contribution of the third asset performance factor to the performance loss of the energy asset over time.

6. The system of claim 1, wherein the first autoencoder model is trained with an unsupervised machine learning algorithm using pairs of images corresponding to energy production data with and without effects associated with the first asset performance factor.

7. The system of claim 1, wherein the instructions further cause the processor to validate the first autoencoder model using real-world energy production data from energy assets.

8. The system of claim 1, wherein the instructions further cause the processor to:generate a clear-sky prediction of the energy asset based on contributions of one or more asset performance factors, including the first asset performance factor, on the performance loss of the energy asset over time; anddetermine a degradation rate of the energy asset based on tracking clear-sky predictions of the energy asset over time.

9. The system of claim 1, wherein the instructions further cause the processor to generate a maintenance and / or service task for the energy asset based on contributions of one or more asset performance factors.Attorney Docket No. 21830-160619-US10. The system of claim 1, wherein the instructions further cause the processor to generate machine-readable command to the energy asset or an associated device to modify an operation of the energy asset.

11. A method for image-based energy asset performance analytics, comprising: transforming time-series energy production data of an energy asset into a first image representing an amount of energy produced by the energy asset as a function of time;generating a first residual image using a first autoencoder model trained to remove effects associated with a first asset performance factor from the first image; and determining a contribution of the first asset performance factor to a performance loss of the energy asset over time based on comparing the first residual image to the first image.

12. The method of claim 11, further comprising: determining contributions of a plurality of asset performance factors using a plurality of autoencoder models, each trained to remove effects associated with a different one of the plurality of asset performance factors.

13. The method of claim 12, wherein the plurality of asset performance factors includes two or more of weather conditions, shading, hardware faults, and / or soiling.

14. The method of claim 11, further comprising: generating a second residual image using a second autoencoder model trained to remove effects associated with a second asset performance factor from the first residual image; and comparing the second residual image to the first residual image to determine a contribution of the second asset performance factor to the performance loss of the energy asset over time.

15. The method of claim 14, further comprising: generating a third residual image using a third autoencoder model trained to remove effects associated with a third asset performance factor from the second residual image; and comparing the third residual image to the secondAttorney Docket No. 21830-160619-US1residual image to determine a contribution of the third asset performance factor to the performance loss of the energy asset over time.

16. The method of claim 11, wherein the first autoencoder model is trained with an unsupervised machine learning algorithm using pairs of images corresponding to energy production data with and without effects associated with the first asset performance factor.

17. The method of claim 11, further comprising: validating the first autoencoder model using real-world energy production data from energy assets.

18. The method of claim 11, further comprising:generating a clear-sky prediction of the energy asset based on contributions of one or more asset performance factors, including the first asset performance factor, on the performance loss of the energy asset over time; anddetermining a degradation rate of the energy asset based on tracking clear-sky predictions of the energy asset over time.

19. The method of claim 11, further comprising: generating a maintenance and / or service task for the energy asset based on contributions of one or more asset performance factors.

20. A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to:transform time-series energy production data of an energy asset into a first image representing an amount of energy produced by the energy asset as a function of time;generate a first residual image using a first autoencoder model trained to remove effects associated with a first asset performance factor from the first image; andAttorney Docket No. 21830-160619-USdetermine a contribution of the first asset performance factor to a performance loss of the energy asset over time based on comparing the first residual image to the first image. Attorney Docket No. 21830-160619-US

Images