跳到正文
实时
Hugging Face Blog·· 7 小时前58

Hugging Face Streamhouse 新增 IBM 时序模型,支持实时预测与异常检测

Time-Series Models in the Streamhouse: A New Frontier and Use Cases

要点速览

Hugging Face Streamhouse 平台新增 IBM 时序模型,支持 8k 上下文窗口、175B 参数规模,可在 Confluent 云端或本地部署,运行于 CPU 无需 GPU,已在 Confluent Cloud 和 Confluent Platform for Apache Flink 上实现原生集成。

正文

Real-time forecasting, anomaly detection, and optimization, powered by IBM time-series models with more than 50 million downloads. Now available natively on Confluent, in the cloud and on premises, running on CPUs.

Every business generates a continuous record of change: a machine reading, a payment, a sale, an order, or a change in inventory. These observations form time series, and they carry signals that can help teams anticipate what happens next.

The challenge is turning those signals into decisions while they still matter. A forecast delivered after a production run, a stockout, or a payment authorization may be accurate, but too late to influence the outcome.

IBM and Confluent are bringing time-series foundation models closer to the data as it flows. With Apache Flink, teams can apply forecasting , anomaly detection and semantic intelligence to streaming data using SQL, making model outputs available to the applications and people responsible for the next action.

Foundation models have changed how organizations work with text, images, and documents. Time-series foundation models extend a similar idea to numerical data by learning general patterns from diverse historical data and then applying that knowledge to new series and tasks. The models can provide a starting point without requiring teams to train a separate model for every series. Depending on the task and deployment, teams can also fine-tune a model or combine its outputs with other data and business logic.

More than 50 million downloads and the family of four models. Click to try the models on Confluent.

Figure 1. More than 50 million downloads, and the family of four models behind them

Now the time series foundation models run natively inside Confluent as built-in functions in the cloud and on premises. In Confluent Cloud they are fully managed, with nothing to deploy (documentation). On premises, the models can run within your own infrastructure using Confluent Platform for Apache Flink.(documentation). The models run on CPUs, without requiring GPUs.

The models can support a range of mission-critical use cases, from helping keep production lines in spec and detecting unusual store activity to improving product-level planning. They can also provide signals to assess card payments and forecast trading volume to help place large orders. These building blocks can be applied across manufacturing, retail, banking, and airlines, bringing insights closer to where decisions are made.

See how time series and real-time context come together to create value:

Streamhouse: real-time data and intelligence for agentic era

Businesses operate continuously. Orders are updated, payments arrive, shipments are delayed, and inventory changes throughout the day. Yet the information behind these events often sits across systems that update at different times or rely on periodic batch processes. When applications and AI agents act on outdated or inconsistent information, they risk making decisions that no longer reflect what is happening.

A Streamhouse addresses this by continuously capturing, processing, governing, and serving current business context. Confluent’s Data Streaming Platform brings together five capabilities—connect, stream, process, govern, and query—to make that context available to applications, analytics, and agents. Apache Flink transforms and enriches data as it flows, while Lightning Tables provide low-latency access to the current state.

The Streamhouse on Confluent

Figure 2. The Streamhouse on Confluent: from existing data sources to current business context that agents, applications, analytics and lakehouses use anywhere.

Time-series models add predictive power to this architecture. Flink can prepare the data, invoke forecasting and anomaly detection, and make the results available alongside the events that produced them. Other models—including large language models, statistical models, and specialized machine learning systems—can contribute where appropriate.

The layers of the Streamhouse: agentic experience, end-to-end solutions, stream-native intelligence and the Streamhouse

Figure 3. The intelligence layer of the Streamhouse: time series models, large language models, and ML and statistical models, with orchestration across them.

This approach can help teams in three ways:

-Scale model development. A pretrained model can provide a common starting point across many time series, reducing the need to develop and maintain a separate model for each one.

-Refresh predictions as data changes. Forecasts can incorporate recent observations as part of a streaming workflow, subject to the model’s requirements and the pipeline’s configuration.

-Put signals closer to decisions. Forecasts and anomalies can be routed to operational applications sooner, where teams can combine them with business context, rules, and human judgment.

-Run compact models on CPUs. The models are small enough to run on standard CPUs alongside the streaming pipeline, in the cloud or on premises, which can reduce the need for dedicated GPU infrastructure and simplify deployment.

05-architecture-dynamic

Figure 4. Stream native time-series intelligence on Confluent

Mission-critical use cases across real-enterprises

The following scenarios illustrate how organizations could combine time-series models, streaming data, and application logic.

Use Case 1 | Production optimization: every tub in spec

An ice cream line’s quality depends on overrun—the amount of air whipped into the mix. Today, the process engineer and operator rely on QA lab measurements taken after the tubs are filled. If a new cream lot or a warm, humid day changes the process, hours of production may already need rework by the time the team sees the result.

To get ahead of the problem, a pretrained TTM model reads SCADA telemetry and QA measurements every 10 minutes. It forecasts the share of tubs expected to be in spec four hours ahead, along with a likely range.

When the forecast falls below the plant’s target, an optimization process evaluates combinations of freezer draw temperature and airflow. Because the model takes these control variables as inputs, it can score each option before anything changes on the line. The operator approves the recommended setpoints, while routine adjustments can be automated within predefined guardrails.

The outcome is a more proactive process: less rework, lower energy use per tonne, and more consistent quality. The same approach can be applied to ovens, dryers, kilns, and mills.

07-production-optimization

Figure 5. A forecast of product quality helps the operator evaluate freezer setpoints before making an adjustment.

Use Case 2 | Anomaly detection: from a flag to the right action

A store operations team monitors sales for thousands of products across hundreds of stores. Fixed thresholds often flag predictable weekend peaks while missing quieter changes, such as a product selling less than expected. As alerts accumulate, it becomes harder for the team to know where to focus.

AI_DETECT_ANOMALIES gives each sales series an expected range based on its recent history. Each new reading is compared with that range at a confidence level the team chooses, without training a separate model for every product and store.

When sales fall outside the expected range, the signal can be combined with inventory and store data to help identify a likely cause and recommend a next step—for example, checking the shelf or replenishing stock. The same approach can be applied to machine sensors, API telemetry, meter readings, and payment volumes.

08-anomaly-detection

Figure 6. Store sales are compared with an expected range. In the view shown, recurring peaks remain within the range and no anomaly is flagged.

Use Case 3 | Supply chain planning: every SKU, every store, every hour

A category planner and a replenishment team need to keep products available across every store without carrying excess inventory. But forecasts may be refreshed only once a week, often focusing on top-selling products, while the long tail is managed with simpler rules. When sales patterns change between planning cycles, one store can run out while another has too much stock.

With AI_FORECAST, a pretrained model can forecast demand across products and stores. In this example, it produces a 28-day forecast with a P10–P90 range, refreshed as new sales data arrives. The planning application combines the forecast with inventory, lead time, and the planner’s chosen service level to calculate a reorder point.

When stock is projected to fall below that point, an agent can draft a purchase order for review. The same approach can support planning for spare parts, staffing, ATM cash, and energy demand.

09-supply-chain-planning-v2

Figure 7. A store-level demand forecast informs a reorder point and a proposed purchase order for oat milk.

Use Case 4 | Fraud and security: every card has its own normal

Fraud teams already rely on sophisticated rules and models to evaluate transactions. One continuing challenge is that a threshold applied across many cardholders may not reflect how each person normally spends.

A time-series signal can add that individual context. In this example, each payment is compared with an expected spending range based on the card’s recent history. A transaction outside that range is passed to the bank’s existing fraud systems as an additional input, alongside other signals and business rules.

The bank can then decide whether to approve the payment, request additional authentication, or investigate further. The decision and the signals behind it can be retained for review.

10-fraud-and-security

Figure 8.A payment outside the card’s expected spending range prompts a step-up authentication in an illustrative fraud workflow.

Use Case 5 | Trading: pace the order with the tape

An execution desk needs to divide a large order into smaller trades while managing its effect on the market. Historical volume curves offer a starting point, but today’s trading activity may look very different from the average.

On each one-minute bar, PatchTST-FM forecasts trading volume for the next 30 minutes, with a P10–P90 range. The execution algorithm uses that forecast to pace the order within the participation limit set by the desk.

If a news event changes market activity, the algorithm can pause and replan. If trading is quieter than expected, it can reduce the size of each slice and signal that the order may take longer to complete. Each forecast and the resulting execution decisions can be retained in the audit trail.

12-trading

Figure 9. Volume forecast informs the proposed execution schedule for a 100,000-share order.

Use Case 6 | Semantic search: find every signal that looks like this

A reliability engineer investigating an unusual sensor pattern may want to know whether similar behavior appeared elsewhere in the plant.

TSPulse can produce representations of time-series windows that support similarity search. In a separately built application, those representations could be indexed and used to retrieve historical patterns across equipment. An engineer could search for shapes such as a spike and recovery, a gradual rise, or a step change, then inspect the associated maintenance records and operating conditions.

This application requires an embedding pipeline, an index, and access to the relevant context. Similarity can help narrow an investigation; it does not establish a root cause by itself.

11-semantic-search

Figure 10. Time-series embeddings can support a search for similar patterns across equipment and other signals.

Industry Specific Applications

The same models and building blocks can support different workflows across industries:

  • Industrial AI: from sensor to setpoint. The plant manager wants every batch in spec and no line stopped by a surprise. SCADA, QA lab and energy meter streams are forecast as they arrive. Production optimization moves the setpoints when quality starts to drift, anomaly detection flags the reading no rule was written for, and semantic search finds every pump that showed the same pattern before a failure. Setpoints, work orders and alerts reach the floor inside the cycle, not at the next shift report.

13-twin-plant

Figure 11. The plant twin, from sensor to setpoint. A new cream lot pulls overrun off target. The forecast reads the line, the agent proposes a new draw temperature and air flow, the operator approves, and the tubs are back in spec. Simulated data.

  • Retail: a store and its supply chain. The store manager wants a full shelf on Saturday without overstock on Monday. Sales, weather and promotion streams are forecast for every product in every store. Supply chain planning sets the order, anomaly detection catches the empty shelf or the missed delivery, and fraud and security holds the suspicious card. Orders, store tasks and holds go out before the shelf runs empty.

14-twin-store

Figure 12. The store twin, a store and its supply chain. A promo is coming for oat milk. The forecast prices in the promo and sets the reorder point, the agent raises a purchase order, the planner approves, and the shelf is full on promo day. Simulated data.

  • Banking: payments and the trading desk. The fraud lead wants to stop card testing without blocking good customers, and the trader wants a large order done without moving the price. Every card and every symbol is scored against its own expected pattern, event by event. Fraud and security holds the odd swipe before authorization, anomaly detection spots the unusual flow, and trading paces each order with the forecast volume.

  • Airlines: hub operations, fleet and network. Operations control wants to see a disruption coming while there is still time to act. Engine, gate and network streams are forecast and checked in real time. Anomaly detection catches the engine reading that drifts, production optimization rebalances gates and crews, and supply chain planning puts the spare part where the aircraft will be.

In every industry, the Streamhouse does the same work: Flink runs the Granite models to forecast, detect and optimize, Streaming Agents act on the live context, and people approve the exceptions.

Explore the models on Confluent to see how time-series intelligence can support real-time decisions.

Try the models on Confluent

Learn more

来源: Hugging Face Blog · huggingface.co