LSTM Cell Prediction

The foundational mechanism that preserves gradient flow across hundreds of time steps. The CEC is a self-connected linear unit with a fixed weight of 1.0 that allows error signals to propagate backward through time without decay. In the context of PRB utilization prediction, this means the network can remember a traffic surge from 200 time steps ago and correlate it with a current congestion event. Without the CEC, standard RNNs suffer from exponential gradient decay, making them incapable of learning long-range dependencies like diurnal traffic patterns that span thousands of 10ms Transmission Time Intervals (TTIs).

A sigmoid-activated neural layer that outputs values between 0 and 1, determining what proportion of the previous cell state to retain or discard. The forget gate examines the current input vector (e.g., latest CQI reports, RRC connection counts) and the previous hidden state to make this decision. When a cell transitions from peak to off-peak hours, the forget gate learns to flush irrelevant high-load memory, preventing stale state information from corrupting future predictions. This is critical for concept drift adaptation in non-stationary network environments.

A two-stage process controlling what new information is stored in the cell state. First, the input gate layer (sigmoid) decides which values to update. Simultaneously, a tanh layer generates a vector of new candidate values that could be added to the state. In multivariate time-series forecasting, this allows the LSTM to selectively incorporate new features—such as a sudden spike in Channel Quality Indicator (CQI) reports—while ignoring transient noise. The element-wise multiplication of these two vectors determines the final information update.

The final gating stage that produces the hidden state vector passed to the next time step and used for prediction. The output gate uses a sigmoid activation to filter the updated cell state, followed by a tanh squashing function to normalize values between -1 and 1. For QoS-aware balancing, this hidden state encodes a compressed representation of the entire historical load context, which a dense regression layer then decodes into specific forecasts—such as predicted Physical Resource Block (PRB) demand for the next 50 TTIs.

The training algorithm that unrolls the LSTM across the specified lookback window and computes gradients through all time steps simultaneously. Unlike standard RNNs where gradients vanish exponentially, the additive nature of the CEC allows gradients to flow as a linear combination of the forget gate outputs. When training on beam-level load data from a 5G massive MIMO array, BPTT with truncated horizons (e.g., 128 time steps) balances computational efficiency with the ability to capture long-range spatial-temporal correlations across multiple antenna beams.

An architectural enhancement where the gating layers receive direct access to the cell state vector. Standard LSTM gates see only the current input and previous hidden state; peephole connections inject the cell state directly into the forget, input, and output gate computations. This allows precise timing-sensitive decisions—for instance, the forget gate can learn to reset precisely when the cell state reaches a threshold representing maximum PRB capacity. This is particularly valuable for inter-cell load shifting where exact timing of handover triggers is critical.

The foundational neural architecture designed for sequential data that LSTM networks extend and improve upon. Standard RNNs process time-series inputs by maintaining a hidden state that captures information from previous time steps, but suffer from the vanishing gradient problem during backpropagation through time.

• Processes sequences one element at a time
• Maintains a hidden state vector as memory
• Struggles to learn dependencies beyond ~10 time steps
• Forms the theoretical basis for LSTM, GRU, and other gated architectures

A streamlined alternative to the LSTM architecture that uses two gates instead of three: the reset gate and the update gate. GRUs merge the cell state and hidden state into a single vector, reducing computational complexity while often achieving comparable performance on smaller datasets.

• Fewer parameters than LSTM (no separate cell state)
• Faster training and inference times
• Update gate controls information retention
• Reset gate determines how much past information to forget
• Often preferred for resource-constrained edge deployments

The training algorithm that unrolls the recurrent network across time steps and applies standard backpropagation to compute gradients. For LSTM cell prediction, BPTT is essential for learning temporal dependencies in network traffic patterns.

• Unrolls the network into a deep feedforward structure
• Computes gradients across the entire sequence length
• Truncated BPTT limits unrolling to a fixed window for efficiency
• The constant error carousel in LSTM prevents gradient decay during this process
• Critical for learning long-range dependencies in PRB utilization data

The LSTM component that decides what information to discard from the cell state. It examines the previous hidden state and current input, outputting a value between 0 and 1 for each number in the cell state via a sigmoid activation.

• A value of 0 means completely forget
• A value of 1 means completely retain
• Enables the network to reset irrelevant context
• Critical for adapting to sudden traffic pattern shifts
• Prevents saturation of the cell state with stale information

The mechanism that controls what new information is stored in the cell state. It operates in two stages: a sigmoid layer decides which values to update, and a tanh layer creates a vector of new candidate values that could be added to the state.

• Selectively incorporates relevant new observations
• Works in tandem with the forget gate
• Prevents irrelevant noise from corrupting the cell state
• Enables the network to capture sudden spikes in cell load
• Essential for accurate PRB utilization forecasting

The final gating mechanism that determines what part of the cell state to expose as the hidden state output. It applies a sigmoid filter to the concatenated previous hidden state and current input, then multiplies it with a tanh-squashed cell state.

• Controls information flow to subsequent layers
• Filters the cell state for the current prediction task
• Enables selective disclosure of long-term memory
• Directly impacts the quality of cell load forecasts
• The hidden state output feeds into the final regression layer