Power converter

Abstract

A power converter includes a regulator, a value providing system arranged to collect at least one operating point of the power converter, and a predictor operative to generate updated regulator parameters (such as one or more supply factors), the updated regulator parameters being implemented by the regulator to generate an output voltage for powering a load. The updated regulator parameters are determined using a process based on at least one collected operating point sample and predictor parameters obtained from a machine learning process.

Description

Technical Field

[0001] Embodiments of the present disclosure generally relate to power converters. Background Art

[0002] Power converters, particularly DC-DC power converters, are widely used due to their high efficiency and small size. Multiphase DC-DC power converters are particularly suitable for providing large currents at low voltages, as required by highly integrated electronic components such as microprocessors, graphics processors, network processors, etc.

[0003] In a conventional manner, a multiphase power converter typically includes several converter branches referred to as phases. Each phase is connected in parallel to provide a corresponding phase current to a common load. Thus, the output current provided to the load by the multiphase converter is the sum of the phase currents. Any electrical power converter can be considered to include at least one phase, and thus includes single-phase power converters and multiphase power converters.

[0004] Each phase of a single-phase power converter or a multiphase converter can be controlled via a proportional-integral-derivative (PID) regulator. Generally, the PID controller controls the operation of switching devices that are arranged to supply charge or current to a storage circuit, i.e., a capacitor or an inductor, and to allow the phase output current to flow from the storage to the load. The k p 、k i and k d coefficients implemented in the PID regulator are selected to produce desired values of the output current and output voltage. Then, it is also known to adjust the corresponding values of the k p 、k i and k d coefficients of each PID regulator in real time according to, for example, the values of the input current and / or input voltage of the converter and the values of the converter output current and / or output voltage.

[0005] Depending on the corresponding power converter design, the regulator for controlling the operation of each phase can be of the PID type just mentioned, but can also be any other type, including only proportional type, integral type, derivative type, any combination such as proportional-integral, integral-derivative, and proportional-derivative, a regulator implementing at least one higher-order component for controlling power conversion, an incremental regulator, an incremental Σ regulator, a differential regulator, etc.

[0006] New generation processors such as CPUs or GPUs have power-saving features that cause the power demand to vary over time in an unpredictable manner. In such a case, the corresponding DC-DC power converter needs to perform well under various load configurations. In particular, such a power converter must meet stable regulation requirements over a wide output range and must also meet specifications regarding transient load configurations, including short transition times and large load steps. Similar requirements apply to power converters used in power supply circuits where the power demand varies randomly over time, such as VR controllers.

[0007] In addition, passive components such as output capacitors and inductors may exhibit significant variations, and these variations also need to be considered in order to optimize the operation of the power converter. Such variations may involve deviations in the target component values due to the manufacturing of each component, or may be due to the aging of each component. However, when adjusting the parameters implemented in the regulator of the power converter (such as the k p 、k i and k d coefficients in the case of a PID regulator), such variations may not be initially known. SUMMARY OF THE INVENTION

[0008] A first aspect of the embodiments herein proposes a power converter configured to convert an input current and an input voltage into an output current and an output voltage. The power converter includes at least one phase and further includes:

[0009] A regulator that operates to generate at least one control signal using at least one regulator parameter implemented in the regulator, the regulator being connected such that the at least one control signal is used by the power converter to effect the conversion;

[0010] A value providing system that is arranged to collect at least one operating point of the power converter, each operating point being associated with an operating moment of the converter and including, on the one hand, measured values for one or more input parameters among the input current, input voltage, phase input current, phase input voltage, and / or one or more output parameters among the output current, output voltage, phase output current, phase output voltage, for the operating moment, and, on the other hand, at least one value of the target output voltage of the power converter, the at least one value being assigned to the moment of the operating point; and

[0011] A predictor that operates to provide a corresponding updated value for each regulator parameter for further implementation by the regulator.

[0012] According to further embodiments herein, the predictor is configured to determine each updated regulator parameter value using a process based on at least one operating point and also based on predictor parameters, the at least one operating point being collected by a value providing system and the predictor parameters being obtained from a machine learning process.

[0013] Thus, embodiments herein include achieving a further level of optimization of the operation of a power converter by adjusting the parameters of a regulator (i.e., the k p (proportional), k i (integral), and k d (derivative) coefficients in the case of a PID regulator), as well as the conversion control signal adjusted by the regulator.

[0014] Since the updated values of the regulator parameters are determined based on measurements of at least one input parameter and / or at least one output parameter and possibly other measurements, the actual values of the passive components involved, as well as the actual conditions of the input power supply of the converter and the converter load, are taken into account for operation optimization. Additionally, when these conditions change over time, the chained operation of the value providing system and the predictor allows for automatic and repeated modification of the regulator parameters to adapt them to the new conditions. In particular, implementing a machine learning process for updating the regulator parameter values allows for improved adaptation of these values over a wide range of operating scenarios of the load.

[0015] Implementing the machine learning process described herein also allows for optimization of the operation of the power converter while taking into account possible variations in the passive component values due to the manufacturing process of the passive components, without having to measure each passive component used.

[0016] The operation of the power converter is also optimized by considering any drift that may occur in the values of the passive components used in the converter or the load (including such drift due to, for example, temperature changes).

[0017] According to further embodiments, values measured for at least one converter temperature can be additionally collected by the value providing system and provided to the predictor, such that the predictor also uses each measured temperature to determine the updated value of each regulator parameter.

[0018] In the case of a multiphase power converter that includes a plurality of phases for providing a total output current and a total output voltage to a load, the total output current and total output voltage being generated by phase output currents and phase output voltages respectively provided by one of the phases, the input parameters for the operating point can include several phase input currents and phase input voltages, and the output parameters for each operating point can include several phase output currents and phase output voltages. In this way, more accurate adaptation of the regulator parameter values to the actual operating conditions of the power converter can be achieved.

[0019] Preferably, the predictor may be adapted to provide an updated value of each regulator parameter based on a plurality of operating points associated with successive operating instances of the converter, wherein the plurality of operating points corresponds to a fixed number of operating points. In other words, the predictor may provide updated regulator parameter values based on a history including a fixed number of operating points. With this improvement, the predictor may optimize the operation of the power converter to a greater extent, particularly by predicting changes in the operating scheme of the load.

[0020] When determining (deriving) regulator parameter values from a plurality of successive operating points, the predictor may implement a recurrent neural network such that whenever the value providing system provides an additional operating point to the predictor, the additional operating point is added to the plurality of just-used operating points in a FIFO queue manner to obtain an updated plurality of operating points for issuing an additional updated value of each regulator parameter.

[0021] One or more of the following additional features may be advantageously implemented, either individually or in combination with several of them:

[0022] In one embodiment, the power converter is a DC-DC power converter or an AC-DC power converter;

[0023] In one embodiment, the regulator is a proportional, integral, and / or derivative (PID)-based regulator, and at least one regulator parameter includes one or more of the k p 、k i and k d coefficients implemented in the regulator;

[0024] In one embodiment, the predictor includes a look-up table for storing labeled training data, and the predictor selects one piece of the labeled training data as the nearest neighbor to at least one operating point;

[0025] In one embodiment, the predictor implements at least one computational step of a regression type in a computational sequence for issuing an updated value of each regulator parameter from at least one operating point;

[0026] In one embodiment, the predictor is arranged to operate in a feed-forward artificial intelligence manner;

[0027] In one embodiment, the predictor is arranged to operate as a neural network, particularly as a single-layer neural network; and

[0028] The predictor is implemented as a neuromorphic chip or implemented in a neuromorphic chip.

[0029] Other embodiments herein include a method for performing power conversion from input current and input voltage to output current and output voltage, the method comprising:

[0030] Using a regulator to generate at least one effective control signal for power conversion, the regulator performing at least one regulator parameter;

[0031] Collecting at least one operating point occurring during DC-DC power conversion, each operating point being related to an operating moment during power conversion and including on the one hand measured values for one or more input parameters among input current, input voltage, phase input current, phase input voltage, and / or one or more output parameters among output current, output voltage, phase output current, phase output voltage for the operating moment, and on the other hand including at least one value of the target output voltage of the power conversion assigned to the moment of the operating point; and

[0032] Using a predictor to provide a corresponding updated value for each regulator parameter, each updated regulator parameter value being predetermined by the regulator for further implementation.

[0033] According to an embodiment herein, each updated regulator parameter value is determined by the predictor using a process based on at least one collected operating point and also based on predictor parameters obtained from a machine learning process.

[0034] According to a further embodiment, the method includes one or more of the following preliminary operations / 1 / to / 3 / performed during the machine learning process:

[0035] / 1 / Collecting labeled training data, the labeled training data including training operating points and corresponding associated values for each regulator parameter;

[0036] / 2 / Using the labeled training data to train a machine learning model of the predictor to obtain predictor parameters to be used by the predictor to infer each new value of each regulator parameter; and

[0037] / 3 / Transmitting the predictor parameters to the predictor.

[0038] Then, the predictor parameters transmitted in step / 3 / are used to operate the power conversion.

[0039] According to a further embodiment, operation / 2 / is performed using computing hardware provided outside the power converter that provides the power conversion. In one embodiment, the computing hardware is disconnected from the power converter such that the power converter performs the power conversion without being connected to the computing hardware anymore.

[0040] Power conversion according to the embodiments herein can be implemented to supply power to any load, such as a load forming part of a data center or server farm. It can be implemented to supply power to a microprocessor, a graphics processor, or a memory bank.

[0041] According to another embodiment, such a microprocessor or graphics processor can itself form part of a data center or server farm powered by a power supply according to the embodiments herein. Alternatively, the power conversion performed according to the embodiments herein is a first power conversion stage for supplying power to a downstream power converter.

[0042] Generally, the power conversion performed according to the embodiments herein is produced using a power converter according to a first inventive aspect, including the listed improvements and preferred embodiments.

[0043] Note that any resources discussed herein (such as predictors, PID regulators, etc.) can include one or more computerized devices, circuits, power converter circuits, etc. to perform and / or support any or all of the method operations disclosed herein. In other words, one or more computerized devices or processors can be programmed and / or configured to operate as explained herein to perform the different embodiments described herein.

[0044] Other embodiments herein include software programs for performing the steps and operations outlined above and disclosed in detail below. One such embodiment includes a computer program product including a non-transitory computer-readable storage medium (i.e., any computer-readable hardware storage medium) encoded with software instructions for subsequent execution. When the instructions are executed in a computerized device (hardware) having a processor, the processor (hardware) is programmed and / or caused to execute the operations disclosed herein. Such an arrangement is typically provided as software, code, instructions, and / or other data (e.g., data structures) arranged or encoded on a non-transitory computer-readable storage medium such as an optical medium (e.g., CD-ROM), a floppy disk, a hard disk, a memory stick, a storage device, etc., or other media such as firmware in one or more ROMs, RAMs, PROMs, etc., or is provided as an application-specific integrated circuit (ASIC), etc. The software or firmware or other such configuration can be installed on the computerized device to cause the computerized device to execute the techniques illustrated herein.

[0045] Thus, the embodiments herein are directed to a method, system, computer program product, etc. that supports the operations discussed herein.

[0046] One embodiment includes a computer-readable storage medium and / or system having stored thereon instructions for providing power conversion. The instructions, when executed by computer processor hardware (such as one or more co-located or separately located processor devices), cause the computer processor hardware to: i) receive a current sample of an operating setting of a power converter; ii) derive a set of power coefficients from the current sample of the operating setting of the power converter, the power coefficients being control responses of a machine learning that are assigned to a set of previous samples of the operating setting of the power converter to maintain an output voltage within a regulated range; and iii) output the set of power coefficients to a regulator.

[0047] For clarity, the order of the above steps is added. Note that any processing steps discussed herein can be performed in any suitable order.

[0048] Other embodiments of the present disclosure include software programs and / or corresponding hardware to perform any method embodiment steps and operations outlined above and disclosed in detail below.

[0049] It should be understood that, as discussed herein, systems, methods, devices, instructions on a computer-readable storage medium, etc. can also be implemented strictly as software programs, firmware, implemented as a hybrid of software, hardware, and / or firmware, or implemented as separate hardware (such as in a processor (hardware or software), in an operating system, or in a software application).

[0050] As discussed herein, the techniques herein are well-suited for more efficient use of wireless services provided to communication devices. However, it should be noted that the embodiments herein are not limited to use in such applications, and the techniques discussed herein are also well-suited for other applications.

[0051] Additionally, note that although each different feature, technique, configuration, etc. herein can be discussed in different places in the present disclosure, it is intended that, where appropriate, each concept can be optionally performed independently of or in combination with each other. Thus, one or more of the inventions described herein can be implemented and viewed in many different ways.

[0052] Additionally, note that the initial discussion of the embodiments herein (“Summary of the Invention”) purposefully does not specify each embodiment and / or incremental novel aspect of the present disclosure or the invention(s) claimed. Instead, this brief description only presents general embodiments and corresponding novelty points relative to conventional techniques. For additional details and / or possible aspects (arrangements) of the invention(s), the reader is directed to the “Detailed Description” section (which is an overview of the embodiments) and the corresponding drawings of the present disclosure, as further discussed below.

[0053] These and other features of the present invention will now be described with reference to the accompanying drawings, which relate to preferred but non-limiting embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a diagram showing elements of an electric power converter according to the present invention;

[0055] Figure 2 is an example diagram showing a calculation sequence implemented by a predictor according to an embodiment herein;

[0056] Figure 3 is an example diagram showing the application of a PID controller and a power factor according to an embodiment herein;

[0057] Figure 4 is an example diagram showing the mapping of the current operating settings of a power converter to appropriate control coefficients to achieve a desired control response according to an embodiment herein;

[0058] Figure 5 is an example diagram showing the mapping of the current operating settings of a power converter to a set of multiple control coefficients and the derivation of control coefficients from multiple sets to achieve a desired control response according to an embodiment herein;

[0059] Figure 6 is an example diagram showing the use of logic to derive control coefficients to control a power converter according to an embodiment herein;

[0060] Figure 7 is a diagram showing an example computer architecture for performing one or more operations according to an embodiment herein; and

[0061] Figure 8 is an example diagram showing a method according to an embodiment herein. DETAILED DESCRIPTION

[0062] For the sake of clarity, components and elements not directly related to the embodiments herein are not described hereinafter, considering that those skilled in the art know how to implement such components and elements.

[0063] For illustrative purposes, but not limited to such types of embodiments, embodiments herein for a DC-DC power converter and a PID type regulator are now described. However, it should be understood that the embodiments herein can be implemented with any type of power converter and any type of regulator for each type of power converter. Other types of regulators that can be alternatively used include proportional regulators, integral regulators, derivative regulators, proportional-integral regulators, integral-derivative regulators, proportional-derivative regulators, regulators that implement at least one higher-order component for controlling power conversion, incremental regulators, incremental Σ regulators, differential regulators, etc. For the present invention, the regulator only needs to implement at least one regulator parameter to issue at least one signal control, and the at least one signal control is used by the power converter to generate the conversion of input voltage and input current to output voltage and output current.

[0064] The DC-DC power converter according to an embodiment herein supplies electrical power to one or more loads such as a computer motherboard, but preferably, particularly supplies electrical power to a processor in a point-of-load configuration. For such a configuration, one power converter is dedicated to one processor and is next to it on a common printed circuit board.

[0065] In a known manner, the converters described herein include one or more phases connected in parallel between the input of the converter and the output that operates to supply power to the load (i.e., the processor to be powered in this example). In one embodiment, each phase may include two switching devices that respectively produce a connected state during the on period and an isolated state during the off period. Each switching device is operated by a control signal, such as a PWM (pulse width modulation) signal or a PFM (pulse frequency modulation) signal issued by a PID regulator. Preferably, all the switching devices of one converter share one PID regulator.

[0066] In a known manner, a conventional PID regulator (controller) implements k p 、k i and k d coefficients for generating a control signal based on the operating parameters of the converter. The k p coefficient is the so-called proportional gain, the k i coefficient is the so-called integral gain, and the k d coefficient is the so-called derivative gain. For this particular case of a PID regulator, the k p 、k i and k d coefficients are the regulator parameters already mentioned in the overview section of the specification.

[0067] During operation, a converter according to an embodiment herein receives an input current and an input voltage (denoted as I input and V input ) from an external DC power supply. The measured values of the input current I input and the input voltage V input can be repeatedly collected, for example, in each n-cycle operation period of the converter, where n is a non-zero fixed integer such as 16, 32, 64, etc.

[0068] The converter converts the input current I input and the input voltage V input into an output current I output and an output voltage V output transmitted to the load. For the same operation moment, the measured values of the output current I output and the output voltage V output can also be collected.

[0069] In the case of a multiphase converter, depending on the converter design, other values can be measured instead of the converter input / output voltage / current just mentioned. These other values can be related to the current supplied to or emitted by one of the phases, respectively called phase input / output current, and denoted as I phase_input or I phase_output . Similarly, the voltage supplied to or generated by one of the phases can also be used, respectively called phase input / output voltage, and denoted as V phase_input or V phase_output . Such phase input / output current / voltage values can also be used in combination with some or all of the converter input values I input and V input as well as the converter output values I output and V output .

[0070] The collection of one or more of these measured values is performed by a so-called value providing system (such as one or more sensors that monitor the operating parameters of the power converter). The value providing system collects the (multiple) measured values related to the same operation moment of the converter into a set of values called an operating point. Each operating point is further completed by the value providing system with a target output voltage, which is also related to the same operation moment as the measured values of the operating point. The PID regulator uses the target output voltage denoted as V target to generate a control signal so that the output voltage V output actually generated by the converter approaches the target output voltage V target . The target output voltage V targetThe successive values allow controlling the variation of the instantaneous output voltage supplied to the load, especially depending on the active or idle periods of the modules inside the load. They also allow controlling the converter output during transient periods between active and idle periods.

[0071] The value providing system transfers each operating point to the predictor, which determines therefrom the values of k p , k i and k d coefficients to be implemented in each PID regulator. The predictor transfers the determined values of k p , k i and k d to the PID regulators of the converter so that each of these PID regulators implements the k p , k i and k d coefficient values relevant to it from the moment they receive them.

[0072] More specifically, as Figure 1 shown, reference numeral 10 denotes a DC-DC power converter; reference numeral 20 denotes a power supply connected to the input of the power converter 10; reference numeral 30 denotes a load powered by the output of the power converter 10.

[0073] In a non-limiting exemplary embodiment, the power supply 20 is of the DC type, and the load can be a microprocessor, a memory, a laptop computer, a smart phone, a tablet computer, an LED bulb, a TV, etc. Each reference numeral 11 denotes a separate phase of the converter, regardless of its number, and each reference numeral 12 denotes a switching device in each phase 11. The internal structure of each phase 11 is not shown in Figure 1 and can be of any type known in the art. For example, it can be of the buck converter type. For the clarity of the drawings, only one switching device 12 is shown for each phase. Other reference numerals are:

[0074] 13: The regulator of the PID type controller in the example considered

[0075] 14: The predictor

[0076] 15: The value providing system (one or more voltage or current sensors), although it is distributed at several locations in the figure

[0077] The PID regulator 13 (PID controller), the predictor 14, and the value providing system 15 together with the phase 11 are part of the DC-DC power converter 10.

[0078] The value providing system 15 may include one or more voltage sensors and / or one or more current sensors, such as common voltage and / or current sensors, for example a DC resistor for sensing current. These sensors may be combined with a sample and hold unit and an analog-to-digital converter to issue measured values V input , I input , V output , I output , V phase_input , I phase_input , V phase_output , I phase_output of at least some of which the measured values correspond to the common operating instants of the converter. Advantageously, the sampling period may be a multiple of the switching period of phase 11, but the sampling period may also be selected according to the application of the converter, for example in order to update the PID parameters fast enough with respect to load variations. The sampling period may also be selected according to the power consumption resulting from each value measurement and the update of each of the k p , k i and k d values.

[0079] V input , I input , V output , I output , V phase_input , I phase_input , V phase_output , I phase_output of at least some of which the measured values and the target output voltage V target are transmitted by the value providing system 15 (the respective sensors) to the PID regulator 13 for operation by the PID regulator 13 in a manner known prior to the present invention.

[0080] According to one embodiment, the (multiple) operating points (i.e., V input , I input , V output , I output and optionally V phase_output and I phase_output of one or more of the (multiple) measured values, and the target output voltage V target ) are transmitted to the predictor 14 for determining the k p , k i and k d coefficient values to be implemented in the PID regulator 13.

[0081] The operation of the predictor 14 is now described.

[0082] Preferably, the predictor 14 includes a FIFO queue (i.e., a data buffer), such as a memory bank, which is used to store a fixed number of operating points related to consecutive operating moments of the converter. For example, an additional operating point is issued by the value providing system 15 at the end of each sampling time. This additional operating point is stored in the entry unit of the FIFO queue, such as a memory bank, and all previously stored operating points are shifted one unit in the queue towards the last storage unit. One of the operating points stored in the last storage unit of the queue is discarded. All or part of the data in the memory bank is used to determine the next values of the k p , k i and k d coefficients. This allows predicting events such as load changes, voltage changes, phase drops, and any possible events by pre-implementing the k p , k i and k d values applicable to such events.

[0083] To predict the values of k p , k i and k d coefficients in a manner suitable for each application, the predictor 14 implements an algorithm known as a machine learning model. Such a machine learning model can run as embedded software within the predictor 14, or directly in hardware, or any combination of both. This allows the same silicon chip to be used for any application of the converter 10. In particular, using a neuromorphic chip that implements a spiking neural network for the predictor 14 enables a very energy-efficient hardware implementation of the machine learning model.

[0084] A simple machine learning model for the predictor 14 includes storing multiple operating points of the power converter 10 within the predictor and the associated values with the k p , k i and k d coefficients. Preferably, a series of consecutive operating points are stored together with the associated values of the k p , k i and k d coefficients.

[0085] Then, each time the value providing system 15 provides a series of actual operating points, an algorithm (such as the nearest neighbor algorithm) determines which series of previously stored operating points in the series of operating points stored (from machine learning) is closest to the series of actual operating points. The difference between the series of actual operating points and any series of stored operating points can be calculated using any criterion known in the art.

[0086] Then, the k p , k i and kd The value of the coefficient is the value associated with the operation point series that is closest in the stored series of operation points. For such an implementation, the stored series of operation points can be compared with k p , k i and k d The associated values of the coefficients are recorded together in a lookup table inside the predictor 14. They constitute the so-called labeled training data and also constitute the predictor parameters for each new set of values of k p , k i and k d values that the predictor 14 uses to infer updates. This implementation of the embodiment in this article is more suitable when the converter 10 has to adapt to a small number of operating scenarios.

[0087] Another possible machine learning model can be based on regression and can use a neural network. This regression-based implementation allows for continuous changes in k p , k i and k d values and thus avoids value jumps, as these jumps can be caused by the aforementioned nearest neighbor implementation. The minimum computational structure to be implemented in the predictor 14 for this regression-based implementation is as Figure 2 shown. It is commonly referred to as a perceptron of the linear classifier type. To obtain the next value of k p , k i and k d for each coefficient among the coefficients to be transmitted to the PID regulator 13, all measured values of at least some of V input , I input , V output , I output and V phase_input , I phase_input , V phase_output , I phase_output in some or all phases, as well as the target output voltage V target of all operation points stored in the FIFO queue memory set are multiplied by predetermined weights and added together and added to a predetermined bias. Then the result of this combination is input as an independent variable into the activation function dedicated to the k p , k i or k d coefficient. The result of the activation function is the next value of this coefficient to be implemented by the PID regulator 13.

[0088] Each computational structure of this type is a feedforward neuron, and a separate neuron is dedicated to each coefficient among k p , k i and k d coefficients. In Figure 2 weights pand bias p are a predetermined weight and bias, respectively, for use in combination with a measurement value related to the k p coefficient and a target output voltage. f p is an activation function for the k p coefficient. Regarding the k i and k d coefficients, similar meanings apply to weights i , bias i , f i and weights d , biase d , f d . Hidden layers can be added within each neuron in a known manner to more clearly determine the k p , k i and k d values with respect to the operating point. The number of hidden neural layers, the number of operating points determined in combination with respect to each k p , k i and k d and the determination frequency will be selected based on the balance between computational workload, prediction accuracy, and characteristics (especially load-related characteristics) applicable to each converter.

[0089] In Figure 2 , n is the number of operating points (samples) involved in determining the k p , k i and k d values each time, i.e., the number of operating points (samples) for the corresponding power supply parameters in each series. For the predictor 14 as described above, n is the length of the FIFO queue memory set. However, thus when n increases and the required storage amount may become significant for a multiphase converter. Then, a method to reduce this storage amount is to directly store at least a portion of the historical information (e.g., the operating points before the last operating point transmitted from the value providing system 15 to the predictor 14) in the neural network instead of entering a FIFO queue such as a memory set. Such a neural network configuration is known in the art as a recurrent neural network. In such a recurrent neural network, long short-term memory may be preferred as they avoid the vanishing or exploding of gradients.

[0090] As described in the overview section of this specification, the weights and biases of all the k p , k i and k d coefficients are predictor parameters. They are provided to the predictor 14 through a preparatory stage called training. This training is preferably performed by computational hardware / software 40 external to the predictor 14 (see Figure 1) is implemented because determining the predictor parameters based on the labeled training data may require considerable computer resources. The computing hardware / software 40 can be provided as a separate computer or can be accessed via the cloud. This configuration of the computing hardware / software 40 for the training phase is advantageous because the computing hardware / software can be shared among a large number of users, thus allowing the implementation of potentially expensive computing devices in a cost-effective manner. Each user can access the computing hardware / software during the initial training phase of the predictor of his power converter, and then his power converter can operate for a long time without using the computing device again.

[0091] The training phase mainly includes the following three steps:

[0092] Form a set of labeled training data, such as each set includes a series of consecutive operating points of the converter and the associated values of k p 、k i and k d coefficients. In this way, each set of labeled training data describes a possible operating sequence of the converter over time, including the instantaneous values of the input and output voltages and currents (optional phase output voltages and currents), and the instantaneous value of the target output voltage. The expected values of k p 、k i and k d coefficients are associated with each series of consecutive operating points. In the prior art, the expected k p 、k i and k d values are called labels. The labeled training data can be advantageously selected in a manner appropriate to the intended application of the power converter 10, particularly with respect to its load 30, in order to obtain optimized operation of the converter in its specific application later;

[0093] Then, the computing device 40 uses one of the known machine learning processes, such as gradient descent (particularly Newton's method or conjugate gradient algorithm), statistical optimization methods (particularly genetic algorithms), or any process implementing backpropagation, etc., to determine the predictor parameters; and

[0094] Transmit the predictor parameters to the predictor 14 for the predictor 14 to later use the predictor parameters to determine k p 、k i and k d values. Transmitting the predictor parameters to the predictor 14 can be performed by value transmission or by writing the corresponding firmware to be implemented within the predictor 14.

[0095] Then, while the converter 10 supplies DC power to the load 30, the operation of the predictor 14 results in the generation of k p 、k i and kd value. The updated k p , k i and k d values are transmitted to the PID controller 13, such that the PID controller 13 switches from the previously implemented k p , k i and k d value set to the updated values.

[0096] Figure 3 is an example diagram showing a PID controller according to an embodiment herein.

[0097] In this example embodiment, the PID controller 13 receives power coefficients (k p , k i and k d ) settings from the predictor 14. The PID controller uses the received coefficients to set (control) the respective gains of each corresponding P, I, D path as shown.

[0098] Figure 4 is an example diagram showing the mapping of the current operating settings of a power converter according to an embodiment herein to appropriate control coefficients to achieve a desired control response.

[0099] As previously described, the power converter 10 includes a plurality of phases 11; the regulator 13 controls the plurality of phases 11 to convert an input voltage into an output voltage.

[0100] In Figure 4 the example embodiment, an instance of the predictor 14-1 (such as hardware and / or software) operates to receive the currently collected samples of the operating settings 210 of the power converter 10. The operating settings 210 are indicated as data sets 410-1, data sets 410-2, data sets 410-3, etc.

[0101] The data set 410-1 (such as data stored in a plurality of FIFO buffers) is a set of first buffered samples of each of a plurality of parameters (such as V input , I input , etc.) obtained at different sampling times.

[0102] The data set 410-2 (such as data stored in a plurality of FIFO buffers) is a set of second buffered samples of each of a plurality of parameters (such as V input , I input , etc.) obtained at different sampling times.

[0103] The data set 410-3 (such as data stored in a plurality of FIFO buffers) is a set of each of a plurality of parameters (such as V input , I inputa set of third cache samples obtained for each parameter (such as etc.) at different sampling times; and so on.

[0104] Thus, the set of each data sample in the previously collected set of data samples (such as data set 410-1, data set 410-2, etc.) includes a plurality of parameters (such as V input , I input , V output , I output , etc.) for the power converter collected over time, and the corresponding sequences of a plurality of data samples for each parameter.

[0105] As further shown, the predictor 14 operates to convert the currently collected samples of the operating setting 210 of the power converter 10 into appropriate control coefficients 120. In one embodiment, the generated control coefficient 120 is a control response of machine learning, and this control response of machine learning is assigned to a pattern of previously stored samples of the operating setting of the power converter 10 indicated by the data set 410.

[0106] In one embodiment, the currently collected samples of the operating setting 210 of the power converter 10 represent the current operating conditions of the power converter 10. The previously stored samples of the operating setting (such as data set 410-1 indicating the first prior operating condition of the power converter 10, data set 410-2 indicating the second prior operating condition of the power converter 10, data set 410-3 indicating the third prior operating condition of the power converter 10, etc.).

[0107] In this exemplary embodiment, based on prior machine learning, each set in different sets of prior detection conditions (operating setting 210) is mapped to a corresponding appropriate control response.

[0108] More specifically, for the conditions of the power converter 10 indicated by the data set 410-1 (such as the monitored voltage / current setting), the control coefficient 120-1 (such as indicating the setting of each coefficient among one or more coefficients k p , k i , k d ) indicates the corresponding appropriate control response for controlling the power converter 10.

[0109] For the conditions of the power converter 10 indicated by the data set 410-2 (such as the setting), the control coefficient 120-2 (such as indicating the setting of each coefficient among one or more coefficients k p , k i and k d ) indicates the corresponding appropriate control response for controlling the power converter 10.

[0110] For conditions (such as settings) of the power converter 10 indicated by the data set 410-3, the control coefficient 120-3 (such as indicating settings for each of one or more coefficients k p 、k i and k d ) indicates the corresponding appropriate control response for controlling the power converter 10.

[0111] For conditions (such as settings) of the power converter 10 indicated by the data set 410-4, the control information 120-4 (such as indicating settings for each of one or more coefficients k p 、k i and k d ) indicates the corresponding appropriate control response for controlling the power converter 10.

[0112] In this exemplary embodiment, it is assumed that the current operating setting 210 of the power converter 10 (for N samples) is most similar / matching to the setting indicated by the data set 410-3. In other words, the currently (most recently) collected samples of the operating setting 210 of the power converter 10 most closely match the pattern of the previously stored samples of the operating setting of the power converter 10. In this case, the predictor 14-1 maps the data set 410-3 to the appropriate control response indicated by the control coefficient 120-3 for it to be selected and applied to the PID controller 13.

[0113] As previously described, in one embodiment, the generated control information 120 (derived from the control coefficient 120-3) indicates the power coefficient settings for the previous operating conditions (associated with the data set 410-3). The settings of one or more PID coefficients in the power converter 10 specified by the control coefficient 120 maintain the output voltage of the power converter 10 within a desired voltage range.

[0114] After generating the control 120 (such as selecting from the control coefficient 120-3), the predictor 14-1 outputs the selected control coefficient 120 to the PID controller 13 or other suitable resources to control multiple phases.

[0115] Thus, in one embodiment, the predictor 14-1 also operates to map the currently collected samples of the operating setting 210 of the power converter 10 to the previously stored samples of the operating setting of the power converter 10 (such as the data set 410-3) to identify and select the appropriate control coefficient 120-3 for the current operating setting 210 of the power supply. As previously described, the previously stored samples of the operating setting indicated by the data set 410-3 are one set in a set of multiple previously stored samples of the operating setting of the power converter (data set 410).

[0116] Figure 5 is an example diagram showing the mapping of the current operating settings of the power converter 10 according to embodiments herein to a set of multiple control coefficients, and deriving a control coefficient from the set of multiple control coefficients to achieve a desired control response.

[0117] In this example embodiment, the predictor 14-1 identifies that the current operating setting 210 most closely matches both the setting specified by the data set 410-3 and the setting specified by the data set 410-4. In this case, the predictor 14-1 applies interpolation and / or extrapolation techniques to derive the control coefficient 120 from the combination of the control coefficient 120-3 and the control coefficient 120-4.

[0118] Figure 6 is an example diagram showing the use of logic to derive control information to control a power converter according to embodiments herein.

[0119] In this example embodiment, similar to Figure 3 , the processing logic of the predictor 14-2 receives the current operating setting 210 of the power converter 10 such as stored in the buffer 610, and derives the control coefficient 120 based on such information.

[0120] The buffer 610-1 stores samples of V input ; the buffer 610-2 stores samples of I input ; the buffer 610-3 stores samples of the V phase output; the buffer 610-4 stores samples of I phase_output ; and so on.

[0121] The control coefficient 120 indicates the settings applied to the regulator 13 in the foregoing manner.

[0122] Figure 7 is an example block diagram of a computer system for implementing any of the operations discussed previously according to embodiments herein.

[0123] Any resources discussed herein (such as the predictor 14, the regulator 13, etc.) can be configured to include computer processor hardware and / or corresponding executable instructions to perform the different operations discussed herein.

[0124] As shown, the computer system 750 of this example includes an interconnect 711 that couples a computer-readable storage medium 712 (such as a non-transitory type of medium (which can be any suitable type of hardware storage medium in which digital information can be stored and retrieved)), a processor 713 (computer processor hardware), an I / O interface 714, and a communication interface 717.

[0125] The I / O interface 714 supports connections to the repository 780 and the input resource 792.

[0126] The computer-readable storage medium 712 can be any hardware storage device, such as a memory, an optical memory, a hard disk drive, a floppy disk, etc. In one embodiment, the computer-readable storage medium 712 stores instructions and / or data.

[0127] As shown, the computer-readable storage medium 712 can be encoded with the communication predictor application 140-1 (e.g., including instructions) to perform any of the operations discussed herein.

[0128] During operation of one embodiment, the processor 713 accesses the computer-readable storage medium 712 by using the interconnect 711 to initiate, run, execute, interpret, or otherwise execute instructions in the predictor application 140-1 stored on the computer-readable storage medium 712. Execution of the predictor application 140-1 results in the predictor process 140-2 to perform any of the operations and / or processes discussed herein.

[0129] Those skilled in the art will understand that the computer system 750 can include other processes and / or software and hardware components, such as an operating system that controls the allocation and use of the hardware resources for executing the communication management application 140-1.

[0130] According to different embodiments, note that the computer system can reside in any one of various types of devices, including but not limited to mobile computers, personal computer systems, wireless devices, wireless access points, base stations, telephone devices, desktop computers, laptop computers, notebooks, netbook computers, mainframe computer systems, palmtop computers, workstations, network computers, application servers, storage devices, consumer electronic devices (such as cameras, portable video cameras, set-top boxes, mobile devices, video game consoles, handheld video game devices), peripheral devices (such as switches, modems, routers, set-top boxes, content management devices, handheld remote control devices, any type of computing or electronic device), etc. The computer system 750 can reside anywhere, or can be included in any suitable resource in any network environment to implement the functions discussed herein.

[0131] Now, the functions supported by different resources will be discussed via Figure 8 the flowcharts in. Note that the steps in the following flowcharts can be executed in any suitable order.

[0132] Figure 8 FIG. 800 is a flowchart showing an example method according to an embodiment. Note that there will be some overlap with the concepts described above.

[0133] In processing operation 810, the predictor 14 receives a current sample of the operating settings 210 of the power converter 10.

[0134] In processing operation 820, the predictor 14 derives a set 120 of power coefficients (such as k p , k i and / or k d ) from the current sample of the operating settings 210 of the power converter 10, the set 120 of power coefficients being a control response of machine learning that is assigned to a corresponding set of previous samples of the operating settings of the power converter 10 to keep the output voltage within a regulated range.

[0135] In processing operation 830 (such as a sub - operation of processing operation 820), the predictor 14 maps the current sample of the operating settings 210 of the power converter 10 to previous samples of the operating settings of the power converter 10 to identify an appropriate control coefficient 120 for keeping the output voltage within a regulated range.

[0136] In processing operation 840 (such as an alternative to sub - operation 830), the predictor 14 inputs the current sample of the operating settings 210 into the processing of the predictor 14, which operation generates a control coefficient 120 based on the received settings 210.

[0137] In processing operation 850, the predictor 14 outputs the control coefficient 120 to the PID controller 13 to control multiple phases of the power converter 10.

[0138] Although the detailed description has focused on predictor embodiments suitable for implementing nearest - neighbor or regression - based machine - learning models, it should be understood that the present invention is not limited to these specific models and that other models may alternatively be used. In particular, any regression variant and any sequence based on a hidden Markov chain may be used.

[0139] It should also be understood that, in addition to DC - DC, the present invention is applicable to any electric power conversion, particularly AC - DC power conversion, but for illustrative purposes, the detailed description focuses on DC - DC power conversion.

[0140] Finally, it should be further understood that the present invention is applicable to any regulator type and is not limited to PID regulators. In each case, the predictor is suitable for providing an updated value of a parameter, as implemented in a regulator for power conversion.

Claims

1. A power converter for converting an input current and an input voltage into an output current and an output voltage, the power converter comprising a plurality of phases and further comprising: A regulator operable to generate at least one control signal using at least one regulator parameter implemented in the regulator, the regulator being connected such that the at least one control signal is used by the power converter to effect the conversion; A value providing system operable to collect at least one operating point of the power converter, each operating point being associated with an operating instance of the power converter and comprising: i) measured values for the operating instance with respect to one or more input parameters among the input current, the input voltage, the phase input current, and the phase input voltage, and / or with respect to one or more output parameters among the output current, the output voltage, the phase output current, and the phase output voltage, and ii) at least one value of a target output voltage of the power converter assigned to the operating instance of the operating point; and A predictor operable to provide a respective updated value of each regulator parameter for further implementation by the regulator, and the predictor being configured to determine each updated regulator parameter value using a process based on at least one operating point collected by the value providing system and also based on predictor parameters obtained from a machine learning process, wherein when a current sample of an operating point of the power converter most closely matches a previously stored sample of an operating point of the power converter, the predictor is operable to map a data set of the current sample to a control coefficient corresponding to the previously stored sample as the respective updated value, or wherein when a current sample of an operating point of the power converter most closely matches two previously stored control coefficients of an operating point of the power converter, the predictor is operable to apply interpolation and / or extrapolation techniques to derive a control coefficient from a combination of the two previously stored control coefficients as the respective updated value.

2. The power converter according to claim 1, wherein the power converter is a DC-DC power converter or an AC-DC power converter.

3. The power converter according to claim 1, wherein the regulator is a proportional, integral, and / or derivative-based regulator, and the at least one regulator parameter includes one or more of the k p , k i , and k d coefficients implemented in the regulator.

4. The power converter according to claim 1, wherein the predictor is operable to determine the updated value of each regulator parameter based on a plurality of operating points associated with consecutive operating instances of the power converter, the plurality of operating points corresponding to a fixed number of operating points.

5. The power converter according to claim 4, wherein the predictor is operable to implement a recurrent neural network such that whenever the value providing system provides an additional operating point to the predictor, the additional operating point is added to the plurality of operating points in a FIFO queue manner to obtain an updated plurality of operating points for issuing an additional updated value of each regulator parameter.

6. The power converter according to claim 1, wherein the predictor includes a look-up table for storing tagged training data and includes computing hardware that is operative to select one piece of training data from the tagged training data as the nearest neighbor to the at least one operating point to determine settings for the PID.

7. The power converter according to claim 1, wherein the predictor includes a computing sequence for emitting the updated value of each regulator parameter from the at least one operating point, and means for implementing at least one computing step of a regression type.

8. The power converter according to claim 1, wherein the predictor is arranged to operate in a feed-forward artificial intelligence manner.

9. The power converter according to claim 1, wherein the predictor is arranged to operate as a neural network.

10. The power converter according to claim 1, wherein the predictor is based on a neuromorphic chip.

11. A method for performing an electric power conversion from input current and input voltage to output current and output voltage using a power converter, the power converter including a plurality of phases and the method comprising: using a regulator to generate at least one control signal for the power conversion, the regulator implementing at least one regulator parameter; collecting at least one operating point occurring during the power conversion, each operating point being associated with an operating moment during the power conversion and including, on the one hand, measured values of one or more input parameters among the input current, the input voltage, the phase input current, the phase input voltage and / or one or more output parameters among the output current, the output voltage, the phase output current, the phase output voltage for the operating moment, and, on the other hand, including at least one value of the target output voltage of the power conversion for the moment assigned to the operating point; and using a predictor to provide a corresponding updated value of each regulator parameter for further implementation by the regulator, wherein each updated regulator parameter value is determined by the predictor using a process based on the at least one collected operating point and also based on predictor parameters obtained from a machine learning process, wherein when the current sample of the operating point of the power converter most closely matches a previously stored sample of the operating point of the power converter, the predictor is operative to map the data set of the current sample to a control coefficient corresponding to the previously stored sample as the corresponding updated value, or wherein when the current sample of the operating point of the power converter most closely matches two previously stored control coefficients of the operating point of the power converter, the predictor is operative to apply interpolation and / or extrapolation techniques to derive a control coefficient from a combination of the two previously stored control coefficients as the corresponding updated value.

12. The method according to claim 11, further comprising operations / 1 / to / 3 / performed during the machine learning process: / 1 / Collect the labeled training data, where the labeled training data includes training operating points and corresponding associated values of each regulator parameter; / 2 / Use the labeled training data to train the machine learning model of the predictor to obtain the predictor parameters to be used by the predictor to infer each new value of each regulator parameter; and / 3 / Transmit the predictor parameters to the predictor, where the power conversion operates using the predictor parameters transmitted in operation / 3 / .

13. The method according to claim 12, wherein operation / 2 / is performed using computing hardware provided outside the power converter for performing the power conversion, and the computing hardware is disconnected from the power converter when the power converter performs the power conversion.

14. The method according to claim 11, wherein the power conversion is used to provide electrical power to a load.

15. An apparatus, comprising: A power converter including a plurality of phases, the plurality of phases operating to convert an input voltage into an output voltage; A regulator operating to control the plurality of phases of the power converter; and A predictor operating to: i) Receive a current sample of the operating settings of the power converter; ii) Derive a set of power coefficients from the current sample of the operating settings of the power converter, the power coefficients being a control response of machine learning, the control response of machine learning being assigned to a set of previous samples of the operating settings of the power converter to maintain the output voltage within a regulated range; and iii) Output the set of power coefficients to the regulator, where when the current sample of the operating settings of the power converter most closely matches one previously stored sample of the operating settings of the power converter, the predictor operates to map the set of data of the current sample to the power coefficients corresponding to the one previously stored sample, or where when the current sample of the operating settings of the power converter most closely matches two previously stored power coefficients of the operating settings of the power converter, the predictor operates to apply interpolation and / or extrapolation techniques to derive the power coefficients from a combination of the two previously stored power coefficients.

16. The apparatus according to claim 15, wherein the current sample of the operating settings of the power converter represents the current operating conditions of the power converter; and where the set of previous samples of the operating settings of the power converter indicates the previous operating conditions of the power converter detected during a machine learning process, the control response of machine learning being assigned to the set of previous samples of the operating settings of the power converter.

17. The apparatus according to claim 16, wherein the set of power coefficients indicates at least one PID coefficient setting to be applied to the regulator, and applying the set of power coefficient settings to the regulator operates to maintain the output voltage within a desired voltage range.

18. The apparatus according to claim 15, wherein the current sample of the operating settings of the power converter comprises a respective sequence of a plurality of data samples for each of a plurality of parameters of the power converter collected over time.

19. The apparatus according to claim 18, wherein the respective sequence of the plurality of data samples comprises: a first cached sample sequence that measures the magnitude of the input voltage at a plurality of sampling times; a second cached sample sequence that measures the magnitude of the input current supplied to the plurality of phases by the input voltage at the plurality of sampling times; a third cached sample sequence that measures the magnitude of the output voltage at the plurality of sampling times; and a fourth cached sample sequence that measures the magnitude of the output current supplied to the load by the output voltage at the plurality of sampling times.

20. The apparatus according to claim 15, wherein the regulator operates at a much higher operating frequency to regulate the magnitude of the output voltage compared to the frequency at which the predictor updates the power factor based on the current sample of the operating settings of the power converter.

21. The apparatus according to claim 15, further comprises: a mode controller operative to power off the predictor for a period of time after the predictor derives a set of power factors, and the mode controller is further operative to power on the predictor again after the period of time.

22. The apparatus according to claim 15, wherein the predictor is further operative to map the current sample of the operating settings of the power converter to a set of previous samples of the operating settings of the power converter to identify a set of power supply factors, the set of previous samples of the operating settings of the power converter being one of a plurality of sets of previous samples of the operating settings of the power converter, and each set in the previous set of samples of the operating settings is assigned a different respective machine learning control response.

23. The apparatus according to claim 15, wherein the current sample of the operating settings of the power converter substantially matches the previous sample of the operating settings of the power converter.

24. The apparatus according to claim 15, wherein the machine learning control response depends on a history of a plurality of sets of previous samples of the operating settings of the power converter.

Images