Neural Network Execution Block With Waveform Generators For Time Series Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning approaches for time series forecasting often rely on off-the-shelf deep learning components without incorporating domain-specific knowledge, leading to suboptimal performance in time series tasks.
Innovation Solution
A novel neural network architecture featuring a fully connected stack of layers with parallel branches, including waveform generator sub-units that constrain waveform generation, enabling specialized function and improved time series forecasting by producing both forecasts and backcasts to form residuals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If off-the-shelf deep learning components (RNN, CNN) are used for time series forecasting, then the system can be implemented with standard architectures, but the performance is suboptimal because domain-specific knowledge is not encoded
Solution Approach 1:
The patent applies local quality by making each execution block specialized for specific waveform patterns (sinusoidal, linear, exponential) rather than using a generic architecture. Each block has dedicated waveform generators that produce specific functional forms, allowing the network to encode domain knowledge about time series patterns locally within each block while maintaining overall system functionality.
Solution Approach 2:
The patent changes parameters by transforming the network output from raw predictions to parameters of predefined waveform functions (amplitude, frequency, phase for sinusoidal; slope and intercept for linear). This parameter transformation allows the network to learn meaningful time series patterns while constraining predictions to physically interpretable forms, improving forecasting accuracy without requiring complex architecture modifications.
2Reliability
If a novel architecture with waveform generators is implemented, then domain-specific knowledge is encoded and forecasting performance improves, but the device complexity increases
Solution Approach 1:
The patent segments the forecasting task into multiple execution blocks, each handling specific waveform patterns. By dividing the overall forecasting function into specialized segments (sinusoidal blocks, linear blocks, exponential blocks), the system can encode domain knowledge in each segment while keeping individual blocks relatively simple and interpretable.
Solution Approach 2:
The execution block design achieves universality by creating a multi-functional unit that can generate multiple types of waveforms (sinusoidal, linear, exponential) through selectable waveform generators. This universal block can handle diverse time series patterns without requiring separate specialized architectures for each pattern type, balancing complexity and functionality.
3Reliability
If fully connected layers with parallel branches are used, then the network can produce both forecasts and backcasts for residual calculation, but the number of layers and parameters increases
Solution Approach 1:
The patent applies preliminary action by computing backcasts (reconstructed past values) in parallel with forecasts using the same execution block. This backcast information is prepared in advance and used to calculate residuals immediately, enabling the network to learn from prediction errors without requiring additional processing steps or layers.
Solution Approach 2:
The patent merges the forecast and backcast computations into a single execution block with parallel branches. By combining these functions in one block rather than using separate networks or sequential processing, the system achieves efficient residual calculation while minimizing the increase in layer count and overall architectural complexity.
Data Source
AI summary
Systems and components for use with neural networks. An execution block and a system architecture using that execution block are disclosed. The execution block uses a fully connected stack of layers and one output is a forecast for a time series while another output is a backcast that can be used to determine a residual from the input to the execution block. The execution block uses a waveform generator sub-unit whose parameters can be judiciously selected to thereby constrain the possible set of waveforms generated. By doing so, the execution block specializes its function. The system using the execution block has been shown to be better than the state of the art in providing solutions to the time series problem.