Streaming data compression and time sequence prediction integrated method for geothermal monitoring platform

Abstract

The invention discloses a streaming data compression and time sequence prediction integrated method for a geothermal monitoring platform, and relates to the technical field of industrial data processing. The method comprises the following steps of: 1, acquiring various original heterogeneous sensing protocol data from a plurality of wellheads and a plurality of layers of sensor arrays in real time on a wellhead gateway side of a geothermal monitoring platform; 2, calling a well layer wavelet domain self-adaptive compression encoder through a well site edge end processor, and performing real-time compression on the structured geothermal flow data; and 3, in a well site edge end processor, performing wavelet dynamic system coding on the multi-scale wavelet coefficient of the well layer wavelet domain compressed code stream, and outputting a temperature prediction sequence, a pressure prediction sequence and a flow velocity prediction sequence of multiple time steps in the future.

Description

Technical Field

[0001] This invention relates to the field of industrial data processing technology, and particularly to the field of geothermal monitoring technology, specifically an integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms. Background Technology

[0002] Geothermal energy, as a renewable and clean energy source, is gradually gaining an important position in the global energy structure due to its high reliability and low life-cycle cost. With advancements in geothermal development technology, especially in the utilization of deep geothermal and hot dry rock resources, the construction of geothermal monitoring systems has become increasingly crucial. A geothermal monitoring platform typically includes a wellhead gateway, sensor array, edge processing unit, and cloud data management center. Its core objective is to acquire accurate physical parameters such as temperature, pressure, and flow rate while ensuring operational safety, reflecting the formation fluid state, wellbore thermal conditions, and system operational stability. Currently, mainstream geothermal monitoring systems generally adopt a timed sampling and batch transmission mode, periodically collecting downhole sensor data and uploading it to a data server via fiber optics, cables, or industrial buses. However, with the increase in monitoring points, the improvement in sensing frequency, and the establishment of multi-parameter linkage models in the well, this traditional acquisition-transmission-storage-analysis model faces numerous challenges, urgently requiring breakthroughs in underlying architecture and data flow processing methods.

[0003] First, in existing geothermal monitoring technologies, sensor data sources are highly heterogeneous. Downhole sensors from different manufacturers often use different communication protocols, such as Modbus, CAN, RS-485, and 4-20mA, which differ in data frame format, timestamp rules, synchronization mechanisms, and data encoding methods. To integrate this data, current systems largely rely on manually configured protocol parsing libraries and hard-coded data structure matching methods. This not only increases the complexity of system integration but also makes it difficult to adapt to the rapid access requirements of adding new equipment or temporarily replacing sensors in the field. Second, in large-scale well cluster systems consisting of multiple wellheads and multiple well layers, the time synchronization mechanism for data from different well locations is inadequate. This easily leads to data out-of-order problems due to sensor time drift or network latency, severely affecting the accuracy of dynamic analysis of well layer conditions and the feasibility of spatial joint modeling.

[0004] Secondly, existing technologies have significant shortcomings in data compression processing. Most mainstream compression methods are based on general-purpose compression algorithms (such as LZW, gzip, bzip2, etc.) or simple time-series prediction differentials (such as Delta encoding). These methods lack the ability to model the cross-features between multiple geothermal physical parameters, and cannot utilize the multi-layered correlations of parameters such as temperature, pressure, and flow velocity in the time, frequency, and scale domains. Consequently, their compression efficiency is low, and their effectiveness in preserving key dynamic change characteristics is limited. Furthermore, these general-purpose algorithms cannot meet the low-latency compression requirements of streaming data scenarios. In situations with high acquisition frequencies (e.g., multiple times per second), they often lead to buffer accumulation, thus affecting the timeliness of data in subsequent links. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms. This method can perform structured processing on heterogeneous protocol data collected by multiple wellheads and multiple layers of sensors at the well site edge. Through wavelet domain adaptive compression and wavelet dynamic system modeling, it achieves efficient compression of streaming data and prediction of temperature, pressure, and flow velocity at multiple future time steps. The entire process retains well location, well depth, and timestamp information, and has beneficial effects such as high compression ratio, strong prediction accuracy, low response delay, and clear spatial positioning, significantly improving the data processing capabilities and intelligence level of geothermal monitoring platforms.

[0006] To address the aforementioned technical problems, this invention provides an integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms. The method includes: Step 1: At the wellhead gateway side of the geothermal monitoring platform, collect real-time raw heterogeneous sensor protocol data from multiple wellheads and multi-layer sensor arrays; unpack and time-align all raw heterogeneous sensor protocol data to form geothermal unified flow format data with a unified structured flow data format; add a unique well location identifier and well depth identifier to each frame of geothermal unified flow format data to obtain structured geothermal flow data; Step 2: The structured geothermal flow data is compressed in real time by calling the well layer wavelet domain adaptive compression encoder through the well site edge processor; Step 3: In the edge processor of the well site, the multi-scale wavelet coefficients of the wavelet domain compressed code stream of the well layer are encoded by wavelet dynamic system, and the wavelet coefficients are directly mapped to the dynamic sequence blocks of the self-evolving state. The dynamic sequence blocks are then time-projected through a folding recursive expansion mechanism. Throughout the projection process, the well location identifier, well depth identifier and timestamp are kept consistent, and the temperature prediction sequence, pressure prediction sequence and flow velocity prediction sequence for future multi-time steps are output.

[0007] Furthermore, in step 1, at the wellhead gateway side of the geothermal monitoring platform, real-time collection of various raw heterogeneous sensing protocol data from multiple wellheads and multi-layer sensor arrays includes at least: temperature data, pressure data, and flow rate data.

[0008] Furthermore, step 2 specifically includes: extracting multi-layer flow segments from the geothermal unified flow format data based on well location and well depth identifiers; sequentially performing wavelet decomposition, coefficient reordering, and threshold pruning on the multi-layer flow segments to form a set of sparse coefficients in the wavelet domain; performing adaptive encoding on the set of sparse coefficients in the wavelet domain according to a fixed sliding window length to generate a compressed code stream in the well layer wavelet domain, and losslessly preserving the well location and well depth identifiers in the compressed code stream.

[0009] Furthermore, the process of extracting multi-layer flow segments from geothermal unified flow format data based on well location and depth identifiers includes: traversing all data frames within the geothermal unified flow format data, comparing the well location identifiers at the end of the data frames, and writing data frames with the same well location identifier into a dedicated well location buffer; stable sorting of data frames in the dedicated well location buffer according to the ascending order of well depth identifiers to generate a depth-ordered data sequence; dividing the data frames into multiple depth groups based on fixed depth intervals within the depth-ordered data sequence, with each depth group corresponding to a well depth interval; establishing a time window queue for each depth group and writing them into the corresponding queue in ascending order of timestamps; checking the continuity of timestamps within each time window queue, and injecting a continuity interruption flag at the jump position if a time jump occurs; and marking the data frames as continuous... The position of the continuity interruption flag is determined by performing time interpolation using the temperature, pressure, and velocity data from the most recent valid data frame, and inserting the interpolation flag. Velocity data is read within each time window queue; if the velocity data is below the well flow lower limit threshold, a zero flow flag is added to the end of the data frame, and it is excluded from subsequent flow segment statistics. For the remaining data frames, the positive and zero velocity data alternation points are detected, and the positive velocity start point and the last data frame before the next zero velocity occurrence are defined together as the start and end boundaries of a flow segment. A flow segment number is generated based on the combination of the well location identifier, well depth identifier, and the timestamps of the start and end boundaries, and all data frames within the same flow segment are written to the flow segment buffer. The flow segment buffer is read sequentially, and each flow segment is written to the input channel of the well wavelet domain compression encoder, forming a complete multi-layer well flow segment set.

[0010] Furthermore, in step 3, the dynamic sequence block is obtained through the following process: using the first frame data of the wellhead wavelet domain compressed code stream as a reference, a wellhead wavelet state reference table is established; phase de-embedding is performed on the wellhead wavelet domain compressed code stream to obtain the wavelet phase mapping matrix; a multi-scale linked list is constructed based on the wavelet phase mapping matrix, with each node of the linked list corresponding to a well depth identifier; fixed-length system windows are divided on the multi-scale linked list according to the timestamp order; wavelet dynamic sequences are extracted within each system window to form a dynamic sequence block.

[0011] Furthermore, in step 3, the process of performing time projection on the dynamic sequence block through the folding recursive expansion mechanism includes: folding and overlapping adjacent dynamic sequence blocks to output a folded and overlapped sequence; performing phase synchronization on the folded and overlapped sequence to avoid phase drift; writing the phase-synchronized folded and overlapped sequence into the wavelet dynamic field buffer to construct a wavelet dynamic field; performing a recursive expansion operation within the wavelet dynamic field to generate a recursive expansion sequence; performing time projection on the recursive expansion sequence to obtain a time-projected sequence; mapping the time-projected sequence back to the well layer wavelet domain according to the well depth identifier to form a well layer wavelet domain prediction coefficient set; and performing inverse wavelet synthesis on the well layer wavelet domain prediction coefficient set to generate future multi-time-step temperature prediction sequences, pressure prediction sequences, and flow velocity prediction sequences.

[0012] Furthermore, in step 3, forward folding pointers and backward folding pointers are set in the dynamic sequence block buffer. Initially, the two pointers point to the first and second data points of the adjacent first and second dynamic sequence blocks, respectively, forming a fixed bidirectional folding pointer pair. Based on the fixed bidirectional folding pointer pair, the dynamic sequence blocks pointed to by the forward folding pointer are added to the folding buffer in the original time order, and then the dynamic sequence blocks pointed to by the backward folding pointer are added to the folding buffer in the reverse time order, completing one bidirectional fold. If there are still remaining dynamic sequence blocks in the folding buffer and their number is even, the folding operation is repeated sequentially until all the dynamic sequence blocks in the folding buffer are arranged in a bidirectional folding manner. The data points in the folding buffer are added and superimposed pair by pair to generate a folded superimposed sequence. When a gap occurs, the value of the previous valid data point remains unchanged to fill the gap, so as to ensure that the length of the folded superimposed sequence is consistent with the total length of the original dynamic sequence block.

[0013] Furthermore, in step 3, the phase identifier of each data point in the folded and superimposed sequence is read and compared with the phase identifier of the first data point in the folded and superimposed sequence. If the difference exceeds the preset phase synchronization threshold, a phase rotation is performed at the corresponding data point position to reset its phase identifier to the first phase identifier, thereby achieving overall phase synchronization and avoiding phase drift.

[0014] Furthermore, the phase-synchronized folded and superimposed sequence is written into the wavelet dynamic field buffer in timestamp order, and synchronization identifier fields are inserted at the beginning and end of the wavelet dynamic field buffer to complete the construction of the wavelet dynamic field. Starting from the first data point in the wavelet dynamic field buffer, data points are read sequentially in the direction of increasing timestamps, and a fixed-step recursive operation is performed: for each data point read, the difference between it and the previous data point is written into the recursive expansion sequence; when the synchronization identifier field is read, the difference accumulator is reset, and the recursion continues until all data points are processed, generating the recursive expansion sequence; the recursive expansion sequence is read and rearranged according to a fixed time projection interval, and the last difference in each time projection interval is written into the time projection sequence; if there is no valid difference in a certain time projection interval, the last difference in the previous time projection interval is copied to fill it, ensuring that the length of the time projection sequence is consistent with the length of the recursive expansion sequence.

[0015] The integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms, as presented in this invention, offers the following advantages: This invention provides an integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms. It enables a unified processing flow for real-time acquisition, structural unification, compression encoding, and future state prediction of heterogeneous data from multiple wellheads and multiple layers of sensors at the well site edge. This significantly improves the overall performance of the geothermal monitoring system in terms of processing efficiency, data quality, prediction capability, and real-time response. By automatically unpacking and aligning the original heterogeneous sensor protocol data at the wellhead gateway side, and assigning unique well location and depth identifiers, the physical source and spatial positioning of each frame of geothermal data are clearly defined, laying the foundation for subsequent well-level processing. The use of a well-level wavelet domain adaptive compression encoder for real-time compression of structured data significantly reduces data redundancy while maintaining key dynamic change characteristics, alleviating the bandwidth burden on the edge and cloud. By constructing wavelet dynamic sequences and employing a folding recursive unfolding mechanism for time projection, multi-time-step prediction is directly completed within the compression domain, avoiding the decompression and reconstruction process and improving prediction efficiency and energy consistency. Meanwhile, this invention retains well location, well depth, and timestamp fields throughout the entire process, ensuring the traceability and consistency of the data stream across the compression and prediction links. This technical solution supports parallel processing of different well layers, possesses high scalability, and can be flexibly deployed in complex well cluster systems. It provides low-latency, high-precision, and high-throughput technical support for geothermal resource status monitoring, early warning analysis, and control strategies, demonstrating significant engineering practical value and promising prospects for industrial application. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the method flow for the integrated method of streaming data compression and time-series prediction for geothermal monitoring platforms provided in an embodiment of the present invention. Figure 2 A schematic diagram of the experimental curves for the wellhead wavelet state reference table and phase de-embedding provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] An integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms, comprising: Step 1: At the wellhead gateway side of the geothermal monitoring platform, collect real-time raw heterogeneous sensor protocol data from multiple wellheads and multi-layer sensor arrays; unpack and time-align all raw heterogeneous sensor protocol data to form geothermal unified flow format data with a unified structured flow data format; add a unique well location identifier and well depth identifier to each frame of geothermal unified flow format data to obtain structured geothermal flow data; In the integrated approach of streaming data compression and time-series prediction for geothermal monitoring platforms, the wellhead gateway, as the first data entry point between each wellhead and the platform's core computing unit, needs to seamlessly access various raw, heterogeneous sensor protocol data generated by multiple wellheads and multi-layer sensor arrays under continuous operation. To achieve real-time acquisition, the wellhead gateway is equipped with a highly reliable industrial-grade acquisition module and a multi-protocol concurrent bus, capable of polling or event-triggered capture of digital messages for core measurements such as temperature, pressure, and flow rate at millisecond-level time resolution. Simultaneously, it maintains local monitoring of sensor heartbeat signals to ensure data link integrity. Due to the significant differences in the operating environments of sensors at different layers within the wellbore, and the varying communication formats used by different vendors, the wellhead gateway immediately invokes a protocol parsing engine to perform protocol unpacking and time-series alignment after data access. The protocol parsing engine incorporates adapters for commonly used serial bus protocols, industrial Ethernet protocols, and customized short-range wireless protocols in existing oil and gas well sites. It quickly locates the payload boundary through feature field identification and a field mapping table, and then uses a message-level verification process to check cyclic redundancy codes, further eliminating noisy or erroneous frames.

[0020] For successfully unpacked valid data frames, the parsing engine reorders messages from different protocol channels based on embedded timestamps and fills in any time gaps with flags, thus completing protocol unpacking and timing alignment. The timing-aligned data is assigned a unified field sequence within the buffer pipeline, including sensor type, original measurement value, sampling timestamp, and device status code, before being written to the geothermal unified stream format data cache. To ensure that the subsequent wellbore wavelet domain adaptive compression encoder can accurately distinguish the source of each well location and depth, the wellhead gateway immediately calls the identifier generator after generating each frame of geothermal unified stream format data to attach a unique well location identifier and well depth identifier to the data frame. The well location identifier uses a platform-wide registered well location coding system to ensure a one-to-one correspondence with existing well site information in the geothermal monitoring platform database; the well depth identifier is mapped according to a depth segment table preset during sensor installation, refined to integer meter level or higher precision, and an error check field prevents identifier mismatch in multi-threaded scenarios. As data frames accumulate in the buffer pool, the wellhead gateway pushes structured geothermal flow data to the edge stream processing channel based on a zero-copy memory mapping mechanism. This data strictly maintains a fixed field order and a unified time base, facilitating direct entry into the wellbore wavelet domain adaptive compression encoder in subsequent steps without further format conversion. Throughout the implementation, the wellhead gateway maintains system time base consistency through dual redundant clock sources and uses a sliding window monitoring algorithm to dynamically balance acquisition latency and buffer depth. If data congestion is detected, a backpressure signal is triggered to notify the upstream sensor array to reduce the sampling frequency, thus achieving smooth data throughput without frame loss. Furthermore, the wellhead gateway employs a local breakpoint resume strategy and incremental verification and recovery mechanism to quickly reconstruct the data stream after brief communication anomalies or power fluctuations, ensuring the reliable mapping relationship between structured geothermal flow data and well location and depth identifiers.

[0021] Step 2: The structured geothermal flow data is compressed in real time by calling the well layer wavelet domain adaptive compression encoder through the well site edge processor; In the integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms, the well site edge processor undertakes the real-time compression task. Internally, it first establishes a high-speed zero-copy receiving channel, writing structured geothermal flow data from the wellhead gateway into a parallel buffer according to system timestamps. Then, it calls the multi-threaded entry point of the well-layer wavelet domain adaptive compression encoder, initiating continuous streaming compression while maintaining a complete mapping between well location and well depth identifiers. After the well-layer wavelet domain adaptive compression encoder starts, it dynamically allocates a wavelet processing pipeline for each data stream and matches the corresponding wavelet reference table based on the well depth identifier. Wavelet decomposition is performed at the first node of the pipeline, simultaneously projecting temperature, pressure, and velocity data into a multi-scale wavelet coefficient space, thus maintaining a unified sampling reference in the time domain and separating detail coefficients and approximation coefficients in the scale domain. The encoder detects the energy distribution of details at each scale using a real-time energy tracking algorithm and establishes an adaptive threshold field. It performs soft threshold reduction on low-energy detail coefficients, progressive preservation on medium-energy detail coefficients, and maintains complete transmission of high-energy detail coefficients, forming a sparse coefficient set.

[0022] Subsequently, the encoder uses a coefficient reordering module to reorder sparse coefficients according to well depth identifier priority and temporal proximity, ensuring that highly correlated coefficients from nearby times within the same well layer are tightly arranged in the byte stream, thereby improving the matching efficiency of the subsequent adaptive variable-length encoder. The adaptive variable-length encoder uses a sliding window length controller to monitor the inflow rate and buffer depth, dynamically adjusting the codeword distribution by statistically analyzing the entropy value of the sparse coefficients within the current window, thus outputting a well layer wavelet domain compressed code stream while maintaining a balance between compression ratio and decoding complexity. The compressed code stream is appended with complete well location and well depth identifiers, as well as necessary time synchronization stamps and coefficient reordering indexes, to avoid mismatches in subsequent prediction links. To cope with fluctuations in field acquisition, the encoder immediately transmits the compression delay index back to the flow control module of the well edge processor after each window completes compression. If the detected delay exceeds the threshold, it triggers a feedback signal downsampling or increases the threshold field intensity to maintain the low-latency characteristics of the compression link. The entire real-time compression process relies on a double-buffered pipeline and a hardware vector acceleration unit to achieve millisecond-level throughput. It provides a breakpoint resume flag and an incremental hash check value at the encoder output to ensure that even under brief network jitter or power interference, the structured geothermal flow data can still flow completely, reliably, and sequentially to the subsequent wavelet dynamic system encoding stage after well-layer wavelet domain adaptive compression encoding. This lays an efficient, scalable, and real-time verifiable data foundation for an integrated method of streaming data compression and time series prediction for geothermal monitoring platforms.

[0023] Step 3: In the edge processor of the well site, the multi-scale wavelet coefficients of the wavelet domain compressed code stream of the well layer are encoded by wavelet dynamic system, and the wavelet coefficients are directly mapped to the dynamic sequence blocks of the self-evolving state. The dynamic sequence blocks are then time-projected through a folding recursive expansion mechanism. Throughout the projection process, the well location identifier, well depth identifier and timestamp are kept consistent, and the temperature prediction sequence, pressure prediction sequence and flow velocity prediction sequence for future multi-time steps are output.

[0024] In the third step of the integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms, the well site edge processor immediately starts the wavelet dynamic system coding engine after receiving the well layer wavelet domain compressed code stream. First, it uses the wavelet phase mapping matrix to de-embed the multi-scale wavelet coefficients, recombining coefficients at different scales but with the same timestamp within the same well layer into a self-cooperative coefficient vector, which serves as the basis for constructing the dynamic sequence block. Subsequently, the coding engine locates nodes in the multi-scale linked list based on the well location and depth identifiers, slides the system window in ascending order of timestamps, and extracts the wavelet dynamic sequence in real time within each system window. Through the built-in self-evolving state generator, the dynamic sequence is mapped into an iterable set of states, forming the first batch of dynamic sequence blocks, which are then synchronously written into the dynamic sequence block buffer. To fully capture the spatial-temporal coupling characteristics of temperature, pressure, and velocity fields within the wellbore, a folding recursive expansion mechanism initiates bidirectional folding pointer pairs within the dynamic sequence block buffer. A folding and overlapping strategy is employed to mirror adjacent dynamic sequence blocks in both forward and reverse directions, creating a highly aligned temporal matching relationship in the folding buffer. Subsequently, the folding buffer is compressed through pairwise addition and overlapping to generate a folded and overlapping sequence. A phase synchronizer is used to detect the phase difference between each data point. When the difference exceeds the phase synchronization threshold, phase rotation is performed to ensure the folded and overlapping sequence remains consistent in the phase domain. After synchronization processing, the folded and overlapping sequence is written into the wavelet dynamic field buffer. Synchronization flag fields are inserted at the beginning and end of the sequence to define the recursion boundaries. During the recursion phase, the recursive expansion operator starts from the first data point in the buffer and reads data points sequentially with a fixed step size. The difference between the current data point and the previous data point is written into the recursive expansion sequence. Upon encountering a synchronization flag field, the difference accumulator is automatically reset to ensure that errors during long-range recursion do not cross the synchronization boundary. After the recursion is completed, the time projection controller performs sequential rearrangement of the recursive unfolded sequence at fixed time projection intervals, writes the last difference value of each time projection interval into the time projection sequence, and at the same time, copies the last difference value of the previous interval to fill in the intervals without valid differences, so as to ensure that the length of the time projection sequence is consistent with that of the recursive unfolded sequence.

[0025] After the time-projected sequence is mapped back to the wellbore wavelet domain through the well depth identifier, a set of predicted coefficients in the wellbore wavelet domain is generated. Then, an inverse wavelet synthesis operation is performed, outputting the predicted temperature, pressure, and flow velocity sequences for future multiple time steps on the time axis corresponding to the current well location and well depth identifiers. The entire wavelet dynamic system encoding and folding recursive expansion mechanism achieves microsecond-level computational granularity with the support of multi-threaded vectorized instructions, and maintains synchronization between the timestamp and the platform's main time base through embedded clock calibration. When input coefficient fluctuations caused by well site environmental disturbances are detected, the gating logic of the dynamic sequence block buffer will immediately adjust the system window length and recursion step size to maintain the stability and continuity of the predicted sequence. Ultimately, the temperature prediction sequence, pressure prediction sequence, and flow velocity prediction sequence are output in their entirety according to the well location identifier, well depth identifier, and timestamp. This not only provides real-time priors for multi-layer thermal-pressure-flow coupling analysis inside the wellbore, but also provides high-precision, multi-time-step prediction inputs for the upper-level cloud decision-making model. As a result, the real-time performance, reliability, and interpretability of the integrated method for stream data compression and time-series prediction for geothermal monitoring platforms are improved as a whole.

[0026] Furthermore, in the integrated approach of streaming data compression and time-series prediction for geothermal monitoring platforms, the real-time acquisition process at the wellhead gateway constructs a unified multi-protocol, multi-wellhead access framework around temperature, pressure, and flow rate data. The platform first installs an industrial-grade protocol bridge at each wellhead, converging raw heterogeneous sensor protocol data output from serial bus, industrial Ethernet, and short-range wireless links to the wellhead gateway via hardware relay. The gateway's built-in high-throughput DMA channel directly drives multiple PHY chips, enabling temperature, pressure, and flow rate data to be completely packaged and written to a shared buffer within milliseconds. To address protocol diversity, the gateway dynamically loads protocol adapter slots based on the sensor array list during initialization and automatically identifies frame header fields using a feature code comparison algorithm. During continuous operation, it performs cyclic redundancy checks and length checks on each data frame, and then calls a time synchronizer to map the embedded device local time to the platform's standard time base, resolving clock drift issues between multiple wellheads and multi-layer sensor arrays. To maintain data order consistency, the gateway uses a timestamp-based priority queue at the caching level to interleave and sort temperature, pressure, and flow rate data from different wellheads. The data is then sent to the protocol unpacking pipeline via a zero-copy exchange mechanism. After unpacking, the field standardization module re-encodes the payload according to a unified stream format template, ensuring that all key fields have a definite positional offset based on byte alignment.

[0027] Subsequently, the identifier generator injects a unique well location identifier and well depth identifier into each frame of data based on the well location coding table and well depth mapping table, and inserts a check stamp to verify the accuracy of the association between the identifier and the data subject. Throughout the process, the gateway continuously monitors the buffer depth and link load. When temperature, pressure, or flow rate data shows a congestion trend, it immediately sends a rate adjustment command to the upstream sensor array to ensure the stability of the acquisition path. With this design, temperature, pressure, and flow rate data can still flow into the subsequent well layer wavelet domain adaptive compression encoder in real time, orderly, and without loss in the form of structured geothermal flow data, even under complex conditions of cross-protocol, cross-wellhead, and cross-depth layers, providing high-quality input for the entire integrated method of streaming data compression and time-series prediction.

[0028] Furthermore, in the integrated method for stream data compression and time-series prediction for geothermal monitoring platforms, the well site edge processor, when processing structured geothermal flow data, first calls the well location index scheduler to perform a full scan of the buffer. A multi-layer well layer mapping table is established based on the secondary key-value pairs of well location and well depth identifiers. Then, based on the increasing timestamp order, fixed depth intervals are divided within the mapping table. Frame sequences of continuous velocity data above the lower flow limit threshold within adjacent depth segments are identified as well layer flow segments. The system generates a unique segment number for each well layer flow segment and writes it to the segment buffer. The segment buffer uses a locked-free ring structure to ensure that simultaneous writing by multiple threads does not generate contention. Subsequently, the wavelet processing scheduling unit pulls velocity, temperature, and pressure data from the segment buffer according to the segment number order and enables multi-scale wavelet decomposition of the pipeline. During wavelet decomposition, the wavelet basis selection module calls the depth-adaptive wavelet reference table based on the well depth identifier, ensuring that higher-order wavelet bases are used to capture local details in areas with drastic changes in thermal gradient, while lower-order wavelet bases are used in stable depth ranges to improve compression efficiency. The decomposed coefficients first enter the coefficient reordering module. This module uses a well-layer local correlation model to perform cross-priority sorting of detail coefficients and approximation coefficients at different scales. High-amplitude detail coefficients with the same well depth and close temporal proximity are placed in the early segment of the byte stream, while low-amplitude detail coefficients are pushed back. This allows the subsequent threshold pruner to identify energy contribution in a single scan. The threshold pruner monitors the difference between the sum of squares of coefficients within the current window and the preset energy retention rate in real time, dynamically adjusting the soft and hard threshold pairs. Coefficients below the soft threshold retain the sign bit and a few significant bits in an exponential decay manner, coefficients in the middle range undergo linear quantization, and coefficients above the hard threshold are fully retained, thus forming a sparse coefficient set in the wavelet domain.

[0029] When the sparse coefficient set is written to the adaptive encoder, the encoder defines the synchronization boundary with reference to a fixed sliding window length. Each window uses a variable-length codeword table and reconstructs the Huffman tree in real time based on the entropy value within the window, minimizing the length of high-probability codewords. Window hashes and synchronization flags are inserted at the beginning and end of the window to ensure continuous decoding across windows. The encoder output wavelet domain compressed bitstream retains well location and well depth flags at the beginning and end, and adds segment number and reordering indexes to ensure lossless reconstruction of the coefficient order of the original well flow segments in any breakpoint recovery scenario. After completing the encoding of each window, the processor immediately feeds back the compression ratio, latency, and frame drop rate to the flow control manager. If the latency approaches the threshold, the flow control manager instructs the threshold pruner to raise the soft threshold to reduce the data volume; conversely, it lowers the soft threshold to improve fidelity. Simultaneously, if the frame drop rate exceeds the limit, the system triggers a well priority weight adjustment, giving the core well flow segments a larger buffer quota. The entire process utilizes an ARM64 vector instruction accelerator and an asynchronous DMA write channel to control the average processing latency of wavelet decomposition, coefficient reordering, and threshold pruning to within a few hundred microseconds per frame, meeting the real-time monitoring requirements of the wellbore. The resulting wellbore wavelet domain compressed codestream, due to its embedded well location and depth identifiers at the codeword level, can be directly mapped back to the flow segment level by the downstream wavelet dynamic system encoding module according to well depth for dynamic sequence construction. This provides a highly sparse, time-consistent, and precisely well-layer-corresponding basic data stream for an integrated method of streaming data compression and time-series prediction for geothermal monitoring platforms.

[0030] Furthermore, the process of generating the well-layer wavelet domain compressed code stream includes: setting a window start pointer at the beginning of the input buffer of the well-layer wavelet domain compressed encoder, and recording the fixed sliding window length as the window span; the window start pointer and the window span together define a continuous coefficient segment as the encoding window; reading all wavelet domain sparse coefficients within the encoding window and writing them into the window coefficient list according to the order of the coefficients in the original time-domain sequence; maintaining a one-to-one correspondence between the window coefficient list and the well location identifier and well depth identifier; using the window coefficient list of the first encoding window as a reference, generating a positional mapping dictionary based on the sorting results of the coefficient occurrence frequency and absolute value; the positional mapping dictionary is only established once in the first encoding window and reused in all subsequent encoding windows, thereby avoiding the redundant overhead of re-counting the dictionary in each window required by the traditional adaptive Huffman method; sequentially searching for continuous coefficients within the window coefficient list. Continue the zero-sequence number segment; when the length of the continuous zero-sequence number segment is greater than 2, write a zero sequence identifier and record the zero sequence length value; for the remaining non-zero coefficients in the window coefficient list, obtain the corresponding codewords by mapping the dictionary according to the bit order; if the codeword length exceeds the preset length threshold, trigger the codeword shortening rule, reduce the codeword length to the threshold, and insert a length extension bit before the codeword to ensure complete codeword decoding; according to the original order of the window coefficient list, write the zero sequence identifier, the zero sequence length value, and the codeword processed by the dynamic codeword length determination sub-step into the encoding output buffer in sequence to form the window encoded string; insert a window synchronization identifier field at the beginning of the window encoded string for subsequent decoding synchronization; write the complete window output string containing the window synchronization identifier field and the window encoded string into the well layer wavelet domain compressed code stream output buffer; then move the window start pointer forward according to the window span.

[0031] Furthermore, in the integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms, when the edge processor at the well site extracts multi-layer flow segments from the unified geothermal streaming data, it first sequentially traverses all data frames within a high-speed circular buffer, reading the well location identifier at the end of each data frame in real time and comparing it with the current well location identifier in the cache. If they match, it is determined that the data frame comes from the same wellhead, and then a zero-copy copying mechanism is used to write the data frame to the corresponding well location-specific buffer. Frames with different well location identifiers are routed to their respective dedicated buffers according to hash indexes, thus forming independent data streams divided by well location in a continuous inflow scenario. Subsequently, each well location-specific buffer starts a depth sorter, calling a stable sorting algorithm to stably sort the data frames according to the well depth identifier from smallest to largest, generating a depth-ordered data sequence. The sorter relies on a double-ended queue structure to ensure that newly arrived frames can be inserted into the correct position in constant time, ensuring that the depth-ordered data sequence maintains global order even in a high-concurrency write environment. The depth segmenter divides the depth-ordered data sequence into multiple depth groups with fixed depth intervals as boundaries. Each depth group corresponds to a well depth interval and a time window queue is established for each depth group. The system time base module writes the data frames into the corresponding time window queue in ascending order of timestamps. If a timestamp jump is detected, a continuity interruption flag is injected at the jump position.

[0032] Upon detecting a continuity interruption flag, the interpolation calculator immediately captures the temperature, pressure, and velocity data from the most recent valid data frame for time interpolation and inserts an interpolation flag at the end of the interpolated frame to ensure the temporal integrity of the time window queue. Within the time window queue, the velocity filter reads velocity data in real time. If the velocity data is below the well flow lower limit threshold, a zero-flow flag is added to the end of the data frame, excluding it from subsequent flow segment statistics. For the remaining data frames, the system calls a positive-zero alternation detector to scan the velocity data sequence, identifying positive-zero velocity alternation points. The positive velocity start point and the last data frame before the next zero velocity occurrence are defined together as the start and end boundaries of a flow segment. The boundary generator generates a unique flow segment number based on the combination of the well location identifier, well depth identifier, and the timestamps of the start and end boundaries, and writes all data frames within the same flow segment into the flow segment buffer. This buffer employs a multi-producer, single-consumer lock-free ring structure to ensure that the order of writing segment numbers during multi-threaded concurrent writing is not disrupted. Finally, the flow segment scheduler reads the contents of the flow segment buffer sequentially and writes each flow segment into the input channel of the well layer wavelet domain compression encoder. Thus, the complete set of flow segments for multiple wellheads and multiple well layers is constructed in real time with millisecond-level latency. This lays the foundation for accurate well location and well depth identification mapping relationships for subsequent wavelet decomposition, coefficient reordering, and threshold pruning. It also further ensures that the overall link of the integrated method for streaming data compression and time series prediction for geothermal monitoring platforms remains stable, traceable, and continuously available in high-throughput scenarios.

[0033] Furthermore, in the third step of the integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms, the well site edge processor immediately invokes the initialization routine upon receiving the first complete time slice of the wellbore wavelet domain compressed code stream. Using the first frame of data in that time slice as a reference, a wellhead wavelet state reference table is established. The wellhead wavelet state reference table adopts a row mapping structure, using the well location identifier and well depth identifier as composite primary keys, corresponding one-to-one with the amplitude, sign bit, and initial phase of wavelet coefficients at each scale under the same timestamp, thereby providing a unified reference system in the subsequent evolution process. Subsequently, the phase de-embedding unit performs phase de-embedding operations on the real-time incoming wellbore wavelet domain compressed code stream. Its core is to use the initial phase recorded in the state reference table to perform point-by-point difference expansion on the current coefficient phase, and to implement phase wrapback when a transition crossing the π-radian threshold is detected, ensuring that the coefficient phase remains continuous on the time axis. After phase de-embedding is completed, the processor constructs a wavelet phase mapping matrix using the de-embedding results. This matrix stores the absolute phase of different scale coefficients in column vector form and the phase distribution of the time series corresponding to different well depths in row vector form. The matrix generation process avoids the copy overhead between multiple threads through shared memory mapping. Next, the linked list generator calls the wavelet phase mapping matrix to create a linked list node for each well depth according to the order of the well depth identifiers from shallow to deep. Each node embeds memory pointers pointing to different scale coefficients at the same timestamp, ultimately constructing a multi-scale linked list. The multi-scale linked list logically presents a top-down hierarchical structure, allowing the processor to focus on the well depth response during vertical traversal and capture scale coupling features during horizontal traversal.

[0034] After the linked list is constructed, the time window scheduler scans the linked list sequentially according to the system clock beat, slicing the entire linked list into fixed-length system windows. The window length is constrained by both the prediction time step and the target computation delay, ensuring that each window contains sufficient historical information without introducing excessive computational bottlenecks. During the window slicing phase, the scheduler inserts synchronization flag fields at the beginning and end of each window for error reset during the recursive phase. Subsequently, the wavelet dynamic sequence extractor enables parallel scanning within each fixed-length system window. It first reads the amplitude and phase pairs of all scale coefficients under the same well depth identifier, aligns the phases across scales using a phase synchronization algorithm, and then calculates the local dynamic increment based on the amplitude change trend and phase precession direction to form the initial wavelet dynamic sequence. To avoid noise amplification, the extractor immediately calls an adaptive smoother after the sequence is formed, weighting the dynamic increment with the weight vector from the wellhead wavelet state reference table. Low-weight coefficients only participate in phase correction without changing the amplitude, ensuring that the energy distribution of the dynamic sequence matches the changes in the physical field. Finally, the smoothed and weighted wavelet dynamic sequences are packaged into dynamic sequence blocks. These blocks retain well location identifiers, well depth identifiers, window start and end timestamps, and cross-scale phase synchronization flags at the data structure level, and append a sequence integrity hash value to the end, facilitating rapid verification of data consistency by the subsequent folding recursive expansion mechanism. Throughout the process, the edge processor utilizes the SIMD accelerated instruction set to vectorize matrix operations and linked list traversals, achieving microsecond-level sequence block generation in conjunction with a zero-copy cache channel. When a drastic jump in wavelet coefficients caused by wellhead environmental oscillations is detected, the system automatically shortens the fixed-length system window to increase phase sampling density and refreshes the timestamp weights of the wellhead wavelet state reference table. This ensures that the dynamic sequence blocks can still accurately capture the transient evolution characteristics of temperature, pressure, and velocity fields, providing high-fidelity, high-consistency spatiotemporal coupling input for the subsequent folding recursive expansion mechanism. This, in turn, improves the overall prediction accuracy and real-time response capability of the integrated streaming data compression and time-series prediction method for geothermal monitoring platforms.

[0035] Furthermore, in the third step of the integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms, the folding recursive unrolling mechanism runs continuously within the well site edge processor. It first activates the folding control unit for the continuously flowing dynamic sequence blocks, pairing adjacent dynamic sequence blocks one by one and writing them into the folding buffer. The folding control unit performs forward and reverse mirror folding using bidirectional folding pointer pairs, performing pairwise addition and superposition operations while maintaining consistency between well location identifiers, well depth identifiers, and timestamps. Empty spaces are filled with the value of the previous valid data point, ultimately outputting the folded superposition sequence. Immediately after the folded superposition sequence is generated, it enters the phase synchronizer. The phase synchronizer reads the phase identifier point by point and compares it with the first and second phase identifiers of the folded superposition sequence. If the difference crosses the phase synchronization threshold, phase rotation is performed, resetting the current point's phase to the first and second phase identifiers, thus avoiding phase drift accumulated during the folding process. The phase-synchronized folded superposition sequence is written to the wavelet dynamic field buffer with zero copies. The buffer inserts synchronization identifier fields at the beginning and end to construct a complete wavelet dynamic field. Next, the recursive expansion operator performs a fixed-step recursive operation on the wavelet dynamic field: starting from the first data point, data points are read sequentially in the direction of increasing timestamp, and the difference between the current data point and the previous data point is written into the recursive expansion sequence; when a synchronization flag field is encountered during the recursion process, the difference accumulator is automatically cleared to zero and the recursion continues until all data points are processed, forming a recursive expansion sequence.

[0036] The time projection controller then performs a sequential rearrangement of the recursive expansion sequence at fixed time projection intervals, writing the last difference value of each time projection interval into the time projection sequence. If a certain interval has no valid difference value, the last difference value of the previous interval is copied to fill it, ensuring that the length of the time projection sequence is consistent with the length of the recursive expansion sequence. After the time projection sequence is completed, the mapping module writes it back to the well layer wavelet domain according to the well depth identifier, generating a set of well layer wavelet domain prediction coefficients. To ensure the physical reversibility of the prediction coefficients, the system appends an integrity hash and a reordering index to the end of the set. The inverse wavelet synthesizer then calls the prediction coefficient set and performs an inverse transformation from the frequency domain to the time domain according to the wavelet reference table used in the well layer wavelet domain compression stage, outputting in parallel the temperature prediction sequence, pressure prediction sequence, and flow velocity prediction sequence for future multiple time steps in the current well location identifier, well depth identifier, and timestamp coordinate system. The entire folding recursive unrolling mechanism achieves millisecond-level data round trips via a SIMD vector instruction accelerator and a shared memory channel. When the system detects folding buffer congestion or wavelet dynamic field drift, it automatically shortens the recursive step size and dynamically adjusts the time projection interval to ensure that the prediction output maintains an optimal balance between real-time performance and accuracy. The final generated temperature, pressure, and velocity prediction sequences not only maintain strict correspondence with the well location, depth, and timestamp of the original well wavelet domain compressed code stream, but also provide accurate and interpretable prior information for upper-level cloud-based geothermal resource management and risk early warning with high spatiotemporal resolution. This, in turn, strengthens the real-time simulation and intelligent decision-making foundation of the integrated streaming data compression and time-series prediction method for geothermal monitoring platforms.

[0037] Furthermore, in the third step of the integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms, the dynamic sequence block buffer is designed as a lock-free double-ended structure with multiple producers and consumers. The buffer's underlying layer uses a circular queue as the main storage container, coupled with an ordered index table based on timestamps, ensuring that the dynamic sequence blocks corresponding to each well location and well depth can be efficiently written in the actual sampling order and read sequentially by the folding recursive unpacking mechanism. When the folding phase is triggered, the folding scheduler first retrieves the first and second data points of the two latest adjacent dynamic sequence blocks in the index table, points the forward folding pointer to the first and second data points of the first dynamic sequence block, and points the backward folding pointer to the first and second data points of the second dynamic sequence block, thus forming a fixed bidirectional folding pointer pair. The folding scheduler then performs a bidirectional fold based on the fixed bidirectional folding pointer pair: the dynamic sequence block pointed to by the forward folding pointer is pushed into the folding buffer point by point in the original time order, occupying the first half of the buffer; the dynamic sequence block pointed to by the backward folding pointer is pushed into the folding buffer point by point in the reverse time order, occupying the second half of the buffer. The two data segments are aligned end-to-end on the time axis, ensuring symmetry at the same time step.

[0038] The folding scheduler checks the remaining space in the folding buffer. If there are still remaining power sequence blocks in the buffer and their number is even, the forward folding pointer and the backward folding pointer are each moved forward by two bits, and the same forward and reverse mirror folding is performed again using adjacent power sequence blocks. This process continues iteratively until all power sequence blocks in the folding buffer are arranged in a bidirectional folding manner. Subsequently, the folding accumulation unit performs a pairwise addition and superposition operation on the data points in the folding buffer: it traverses the corresponding elements of the buffer index i and index len-i-1, adds the values ​​of the two elements, and writes them back to the low address of the folding buffer; if the current folding buffer length is odd, resulting in empty spaces, the system keeps the value of the previous valid data point unchanged to fill the empty spaces, ensuring that the length of the folded superposition sequence is completely consistent with the total length of the original power sequence blocks. The folding accumulation unit is processed in parallel by a SIMD gradient register array at the hardware level, which can complete the addition and superposition of multiple pairs of data points in a single clock cycle, and realizes real-time dequeueing of the folding buffer through a pipelined cascading mechanism. After the folded and stacked sequence is generated, the complete well location identifier, well depth identifier, and window start and end timestamps are automatically migrated to the beginning of the sequence to ensure consistency with the input format of subsequent phase synchronizers. Simultaneously, the folding accumulation unit appends a sequence integrity check code to the end of the sequence to support breakpoint resumption and data consistency verification. The entire fixed bidirectional folding process operates through a time-driven microtask scheduling framework and a high-speed memory-mapped I / O channel. The folding processing latency is stably controlled at the microsecond level, and all write operations are completed locally on the NUMA node, avoiding additional latency introduced by cross-node inconsistency access. This design not only fully utilizes the time inversion symmetry between dynamic sequence blocks to effectively compress redundant information within the same period, but also provides highly aligned input sequences for phase synchronization, recursive unrolling, and time projection, thereby improving the overall computational efficiency and prediction stability of the integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms.

[0039] Furthermore, the phase synchronizer first locks the first data point of the folded and stacked sequence, writes its phase identifier into the local register as the reference phase, and then starts a point-by-point scanning thread to sequentially read the phase identifier of each data point from left to right and compare it with the reference phase. An embedded circular difference accumulator is used in the scanning thread to calculate the difference in real time in continuous data stream scenarios and compare it with a preset phase synchronization threshold. When the difference is less than the phase synchronization threshold, the phase identifier of the current data point is considered to be in sync with the reference phase and requires no adjustment; the scanning thread simply records a pass mark in the register and continues forward. When the difference exceeds the phase synchronization threshold, the system considers that the data point has experienced phase drift and immediately calls the phase rotation operator to perform phase rotation on the data point. The phase rotation operator is based on a vectorized lookup table mechanism, does not perform trigonometric operations, and directly resets the phase identifier of the current data point to the reference phase, completing the repositioning with a nanosecond delay. The correction result is then written back to the corresponding address in the folded and stacked sequence, and a single-bit rotation flag is inserted at the end of the data point for subsequent recursive unfolding stages to identify that the phase has been adjusted. Throughout the scanning process, the phase synchronizer uses dual-channel DMA to load the folded and superimposed sequence into on-chip cache in segments. While performing high-bandwidth read and write operations, it keeps the well location identifier, well depth identifier, and timestamp absolutely unchanged, ensuring that the synchronization operation only affects the phase identifier field and does not touch other service fields.

[0040] To prevent synchronization jitter caused by continuous phase jumps in high-noise wellbore environments, the phase synchronizer incorporates an adaptive threshold tuning module. This module monitors the number of phase rotations and the energy distribution of the folded and superimposed sequence over several past windows in real time. When the number of rotations exceeds the threshold or the energy concentration decreases within a short period, the tuning module automatically raises the preset phase synchronization threshold to suppress oversensitivity to weak noise. Conversely, when the number of rotations is too low or the energy concentration increases over a long period, the tuning module lowers the threshold to improve the ability to detect potential phase drift. To avoid data contention between the synchronization thread and other computing threads, the phase synchronizer employs a read-write separation lock scheme: the scanning thread sequentially traverses using read locks, only briefly upgrading to write locks when a phase rotation is truly necessary. The write lock scope is strictly limited to a single data point, and the entire upgrade cycle is less than one clock cycle. After synchronization, the phase synchronizer performs tail verification on the folded and superimposed sequence. By accumulating the rotation flag and comparing it with the sequence integrity hash, it confirms that all data points requiring correction have been processed and no new byte misalignments have been introduced. Then, the phase-synchronized folded and superimposed sequence is transferred to the wavelet dynamic field buffer for subsequent recursive expansion. This design not only eliminates cross-scale interference caused by phase drift within microsecond-level time slices, ensuring that temperature, pressure, and flow velocity signals remain coherent in the frequency domain, but also guarantees that the integrated method of streaming data compression and time series prediction for geothermal monitoring platforms still has stable and high-precision prediction capabilities in long-term unattended operation scenarios, thus providing reliable support for real-time control and safety early warning of geothermal resource development.

[0041] Furthermore, after the wavelet dynamic field is constructed, the recursive controller starts from the first data point in the buffer and performs a fixed-step recursive operation in the direction of increasing timestamps: for each data point read, the difference between its value and the previous data point is calculated and written into the recursive expansion sequence; if the current data point is a synchronization identifier field, the difference accumulator is immediately cleared to prevent the error from spreading across the boundaries between windows, while the recursive process continues uninterrupted, recursively applying it to subsequent data points until all valid data points in the wavelet dynamic field buffer have been traversed, and finally a complete recursive expansion sequence is output. The time projection controller then reads the recursive expansion sequence and performs sequential rearrangement according to a fixed time projection interval: it slices the sequence by interval, extracts the last difference value in each time projection interval and writes it into the time projection sequence; if no valid difference value is detected in a certain time projection interval, the last difference value of the previous time projection interval is copied to fill the gap, ensuring that the length of the time projection sequence is consistent with the recursive expansion sequence and avoiding time axis defects. To maintain high throughput computing, both the recursive controller and the time projection controller run on SIMD vector units. Difference calculation and interval slicing are fully implemented in the register-level pipeline, while the detection of the synchronization flag field is completed by a hardware comparator within the same clock cycle. The difference accumulator clearing operation is triggered synchronously with the pipeline refresh. The system also deploys a dynamic water level detector within the wavelet dynamic field buffer to monitor write depth and read progress in real time. When the write speed exceeds the read threshold, the system proactively shortens the fixed step size or reduces the time projection interval to prevent buffer overflow; conversely, when the read speed exceeds the write speed, the step size is widened and the projection interval is extended to ensure efficient utilization of computing resources. While keeping the well location identifier, well depth identifier, and timestamp fields absolutely unchanged, the entire process only performs efficient recursion and sequential rearrangement of the amplitude and difference domain data. This not only eliminates the cumulative error of long-term series, but also provides a temporally continuous and energy-conserving difference input for subsequent well layer wavelet domain inverse synthesis. This further enhances the comprehensive performance of the integrated method for streaming data compression and time series prediction for geothermal monitoring platforms in terms of real-time prediction accuracy, link stability, and cross-scale coupling analysis.

[0042] The following example fully demonstrates the integrated method for streaming data compression and time-series prediction for a geothermal monitoring platform in a single-wellhead WELL-A01 scenario with a well depth of D-1500 meters. The example assumes the platform starts data acquisition at midnight Beijing time, running at a fixed sampling frequency of once per second for sixty seconds for the temperature, pressure, and flow velocity channels. For brevity, only the temperature channel's full values ​​and calculations are provided; the pressure and flow velocity calculations can use the exact same computational path. Sixty consecutive temperature samples are collected at the wellhead gateway. Indicates the sampling index, with values... ; Indicates the first The raw temperature measurement in seconds, in degrees Celsius. This example uses the first sixteen sample points to illustrate all subsequent mathematical steps: The well site edge processor calls the well formation wavelet domain adaptive compression encoder to adjust the length. The sequence is subjected to three-level discrete wavelet decomposition. Let... This is the scale layer number. ; The translation index; For the mother wavelet of “db4” at scale With translation Basis functions; For the corresponding wavelet coefficients, the discrete wavelet transform formula is written as: .

[0043] Using the one-dimensional Haar approach (db4 approximation) to calculate the example data, we can obtain the first-order approximation coefficient vector. With first-level detail coefficient vector Continue with Iterative decomposition yields The third iteration produces Therefore, the total number of coefficients remains sixteen, ensuring complete reversibility.

[0044] For adaptive noise reduction, the set of absolute values ​​of all detail coefficients is first denoted as... Take its absolute median. .make , For robust estimation of detail noise, numerical Set a soft threshold. .when At this point, the detail coefficient is set to zero. It is retained after thresholding. One non-zero detail coefficient and all The coarsest approximation coefficient, plus... There are 1 first-order approximation coefficients, totaling 1 Item; the original sequence needs Each wavelet coefficient is still stored using 16-bit fixed-point storage. Compression ratio .

[0045] The encoder then processed this Perform Huffman variable-length coding on each non-zero coefficient, and insert the well location identifier WELL-A01 and the well depth identifier D-1500 at the beginning and end of the code stream to obtain the well layer wavelet domain compressed code stream. Entering the wavelet dynamic system coding stage, first using... The first frame establishes the wellhead wavelet state reference table, and This is denoted as the phase identifier of the first frame. Phase de-embedding is then performed on the subsequent coefficient sequence to obtain the wavelet phase mapping matrix. .set up To fix the system window length, For the first The system window aggregates wavelet dynamic sequence blocks across all scales. Four windows can be constructed using the sixteen samples mentioned earlier: , , , ; Folding control unit For the first pair For the second pair of mirror folding and overlapping: ,in This is the time offset index within the window. .Pick For example, we can obtain: .

[0046] The phase synchronizer reads the phase of each point in the folded and superimposed sequence. .like Then perform a phase rotation and reset. In this example, a phase synchronization threshold is set. After synchronization and Write the wavelet dynamic field buffer according to the timestamp, and insert a special synchronization identifier "SYNC" at the beginning and end of the wavelet dynamic field.

[0047] The recursive controller starts from the beginning of the buffer with a fixed step size. Recursion: ,in It is a recursive difference sequence. This represents the dynamic field sequence value. The command is executed upon reading "SYNC". Then clear the difference accumulator. After all iterations are complete, the recursive expansion sequence is obtained. The time projection controller sets the time projection interval. .make If the time projection interval index is used, then the time projection sample If the first If the interval has no valid difference, then use Replacement. The resulting time-projected sequence. Mapping back to the well layer wavelet domain based on well depth identifiers to generate predicted wavelet coefficients. The inverse wavelet synthesizer uses the same mother wavelet as the compression stage. ,in Indicates the first Temperature forecast per second. Assuming the average of the last four recursive differences... The temperature prediction result for the next five seconds is: .

[0048] Figure 2 This paper demonstrates the complete technical process and experimental verification results of the wellhead wavelet state reference table and phase de-embedding. The figure uses wavelet phase (rad) as the horizontal axis and wavelet phase mapping matrix amplitude as the vertical axis, detailing the entire process from establishing the wellhead wavelet state reference table to extracting the wavelet dynamic sequence. Specifically, this invention first uses the first frame of the wellbore wavelet domain compressed code stream as a reference to establish the wellhead wavelet state reference table. (See figure...) Figure 2 As shown, the first frame of data is referenced at phase 0, providing a stable reference point for the entire phase de-embedding process. The wellhead wavelet state reference table curve exhibits a typical sine wave shape, with an amplitude range from -4.0 to 8.0 and a phase range covering the complete 2π cycle, reflecting the continuous distribution characteristics of wavelet coefficients in the phase domain. Phase de-embedding is performed on the wellbore wavelet domain compressed codestream to obtain the wavelet phase mapping matrix. This matrix reorganizes the wavelet coefficients in the compressed codestream according to phase characteristics, forming a data structure with clear phase identifiers. The phase de-embedding process is achieved through hollow circle-marked feature points, which are distributed at key phase positions such as π / 8, π / 4, 3π / 8, and π / 2, ensuring the accuracy and completeness of the phase mapping. A multi-scale linked list is constructed based on the wavelet phase mapping matrix, with each node corresponding to a well depth identifier. Figure 2 As shown, the well depths are identified as 50m, 100m, and 150m, corresponding to wavelet features at different depths. A multi-scale linked list connects adjacent nodes with dashed lines, forming a hierarchical data organization. Each node contains wavelet coefficient information for its corresponding well depth. Fixed-length system windows are divided on the multi-scale linked list according to timestamp order. Figure 2 The diagram displays four fixed-length system windows, each covering a specific phase range, enabling time-series segmentation of wavelet data. The system windows are divided according to an equal-length principle to ensure a balanced distribution of data within each window. Wavelet dynamic sequences are extracted from each system window to form dynamic sequence blocks. The wavelet dynamic sequence extraction process is indicated by thick black lines, corresponding to the processing results of the four system windows. The extraction process considers the time-domain and frequency-domain characteristics of the wavelet coefficients, generating sequence data with dynamic characteristics through comprehensive analysis of phase and amplitude information. The entire phase de-embedding process maintains the consistency of the well location identifier W05, ensuring the traceability of data processing. The establishment of key data structures such as the wavelet phase mapping matrix and multi-scale linked list nodes lays a solid foundation for subsequent folding recursive unfolding and time projection processing, demonstrating the technical advantages of this invention in geothermal monitoring data processing.

[0049] The present invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make various improvements and modifications to the present invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. An integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms, characterized in that, The method includes: Step 1: At the wellhead gateway side of the geothermal monitoring platform, collect real-time raw heterogeneous sensor protocol data from multiple wellheads and multi-layer sensor arrays; unpack and time-align all raw heterogeneous sensor protocol data to form geothermal unified flow format data with a unified structured flow data format; add a unique well location identifier and well depth identifier to each frame of geothermal unified flow format data to obtain structured geothermal flow data; Step 2: The structured geothermal flow data is compressed in real time by calling the well layer wavelet domain adaptive compression encoder through the well site edge processor; Step 3: In the edge processor of the well site, the multi-scale wavelet coefficients of the wavelet domain compressed code stream of the well layer are encoded by wavelet dynamic system, and the wavelet coefficients are directly mapped to the dynamic sequence blocks of the self-evolving state. The dynamic sequence blocks are then time-projected through a folding recursive expansion mechanism. Throughout the projection process, the well location identifier, well depth identifier and timestamp are kept consistent, and the temperature prediction sequence, pressure prediction sequence and flow velocity prediction sequence for future multi-time steps are output.

2. The integrated method for streaming data compression and time-series prediction for geothermal monitoring platforms as described in claim 1, characterized in that, In step 1, at the wellhead gateway side of the geothermal monitoring platform, real-time collection of various raw heterogeneous sensing protocol data from multiple wellheads and multi-layer sensor arrays includes at least: temperature data, pressure data, and flow rate data.

3. The integrated method for streaming data compression and time-series prediction to a geothermal monitoring platform as described in claim 2, characterized in that, Step 2 specifically includes: extracting multi-layer flow segments from geothermal unified flow format data based on well location and well depth identifiers; sequentially performing wavelet decomposition, coefficient reordering, and threshold pruning on the multi-layer flow segments to form a set of sparse coefficients in the wavelet domain; performing adaptive encoding on the set of sparse coefficients in the wavelet domain according to a fixed sliding window length to generate a compressed code stream in the well layer wavelet domain, and losslessly preserving the well location and well depth identifiers in the compressed code stream.

4. The integrated method for streaming data compression and time-series prediction to a geothermal monitoring platform as described in claim 3, characterized in that, The process of extracting multi-layer flow segments from geothermal unified flow format data based on well location and depth identifiers includes: traversing all data frames within the geothermal unified flow format data, comparing the well location identifiers at the end of each data frame, and writing data frames with the same well location identifier into a dedicated well location buffer; stable sorting of data frames in the dedicated well location buffer according to ascending well depth identifiers to generate a depth-ordered data sequence; dividing the data frames into multiple depth groups based on fixed depth intervals within the depth-ordered data sequence, with each depth group corresponding to a well depth interval; establishing a time window queue for each depth group and writing the data frames into the corresponding queue in ascending order of timestamps; checking the continuity of timestamps within each time window queue, and injecting a continuity interruption flag at the jump position if a time jump occurs; and processing data frames marked as continuous. The location of the break flag is determined by performing time interpolation using the temperature, pressure, and velocity data from the most recent valid data frame, and inserting an interpolation flag. Velocity data is read within each time window queue; if the velocity data is below the well flow lower limit threshold, a zero-flow flag is added to the end of the data frame and it is excluded from subsequent flow segment statistics. For the remaining data frames, the positive-zero alternation points of the velocity data are detected, and the positive velocity start point and the last data frame before the next zero velocity occurrence are defined together as the start and end boundaries of a flow segment. A flow segment number is generated based on the combination of the well location identifier, well depth identifier, and the timestamps of the start and end boundaries, and all data frames within the same flow segment are written to the flow segment buffer. The flow segment buffer is read sequentially, and each flow segment is written to the input channel of the well wavelet domain compression encoder to form a complete set of multi-layer well flow segments.

5. The integrated method for streaming data compression and time-series prediction to a geothermal monitoring platform as described in claim 4, characterized in that, In step 3, the dynamic sequence block is obtained through the following process: using the first frame of the wellhead wavelet domain compressed code stream as a reference, a wellhead wavelet state reference table is established; phase de-embedding is performed on the wellhead wavelet domain compressed code stream to obtain the wavelet phase mapping matrix; a multi-scale linked list is constructed based on the wavelet phase mapping matrix, with each node of the linked list corresponding to a well depth identifier; fixed-length system windows are divided on the multi-scale linked list according to the timestamp order; wavelet dynamic sequences are extracted within each system window to form a dynamic sequence block.

6. The integrated method for streaming data compression and time-series prediction to a geothermal monitoring platform as described in claim 5, characterized in that, In step 3, the process of performing time projection on the dynamic sequence blocks through the folding recursive expansion mechanism includes: folding and overlapping adjacent dynamic sequence blocks to output a folded and overlapped sequence; performing phase synchronization on the folded and overlapped sequence to avoid phase drift; writing the phase-synchronized folded and overlapped sequence into the wavelet dynamic field buffer to construct a wavelet dynamic field; performing a recursive expansion operation within the wavelet dynamic field to generate a recursive expansion sequence; performing time projection on the recursive expansion sequence to obtain a time-projected sequence; mapping the time-projected sequence back to the well layer wavelet domain according to the well depth identifier to form a well layer wavelet domain prediction coefficient set; and performing inverse wavelet synthesis on the well layer wavelet domain prediction coefficient set to generate future multi-time-step temperature prediction sequences, pressure prediction sequences, and flow velocity prediction sequences.

7. The integrated method for streaming data compression and time-series prediction to a geothermal monitoring platform as described in claim 6, characterized in that, In step 3, forward folding pointers and backward folding pointers are set in the dynamic sequence block buffer. Initially, the two pointers point to the first and second data points of the adjacent first dynamic sequence block and second dynamic sequence block, respectively, forming a fixed bidirectional folding pointer pair. Based on a fixed bidirectional folding pointer pair, the power sequence blocks pointed to by the forward folding pointer are added to the folding buffer in their original time order. Then, the power sequence blocks pointed to by the backward folding pointer are added to the folding buffer in reverse time order, thus completing one bidirectional fold. If there are still remaining power sequence blocks in the folding buffer and their number is even, the folding operation is repeated sequentially until all the power sequence blocks in the folding buffer are arranged in a bidirectional folding manner. The data points in the folded buffer are added and superimposed in pairs to generate a folded superimposed sequence. When a gap occurs, the value of the previous valid data point is kept unchanged to fill the gap, so as to ensure that the length of the folded superimposed sequence is consistent with the total length of the original power sequence block.

8. The integrated method for streaming data compression and time-series prediction to a geothermal monitoring platform as described in claim 7, characterized in that, In step 3, the phase identifier of each data point in the folded and superimposed sequence is read and compared with the phase identifier of the first data point in the folded and superimposed sequence. If the difference exceeds the preset phase synchronization threshold, phase rotation is performed at the corresponding data point position to reset its phase identifier to the first phase identifier, thereby achieving overall phase synchronization and avoiding phase drift.

9. The integrated method for streaming data compression and time-series prediction to a geothermal monitoring platform as described in claim 8, characterized in that, The phase-synchronized folded and superimposed sequence is written into the wavelet dynamic field buffer in timestamp order, and synchronization identifier fields are inserted at the beginning and end of the wavelet dynamic field buffer to complete the construction of the wavelet dynamic field. Starting from the first data point in the wavelet dynamic field buffer, data points are read sequentially in the direction of increasing timestamp, and a fixed-step recursive operation is performed: for each data point read, the difference between it and the previous data point is written into the recursive expansion sequence; when the synchronization identifier field is read, the difference accumulator is reset, and the recursion continues until all data points are processed, generating the recursive expansion sequence; the recursive expansion sequence is read and rearranged according to the fixed time projection interval, and the last difference in each time projection interval is written into the time projection sequence; if there is no valid difference in a certain time projection interval, the last difference in the previous time projection interval is copied to fill it, ensuring that the length of the time projection sequence is consistent with the length of the recursive expansion sequence.

Images